> Intent understanding: GPT-5.6 can better infer the user’s underlying goal and intended level of work without you specifying every step. Continue to state important constraints, approval boundaries, and success criteria explicitly.
> Original image detail: GPT-5.6 preserves the original dimensions of images sent with original or auto detail instead of resizing them to a patch budget or pixel-dimension limit.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Control warmth: GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
I don't follow. Isn't "the model actually cares and will do what you say" a reason to use those kinds of instructions more liberally?
That part is confusing because it's not like they provide an example of how default GPT-5.6 output compares with GPT-5.5 both with default output and prompted for brevity. Whenever I use such prompts, it's usually because I want the model to give me the gist in a few sentences. I'd be stunned if GPT-5.6 was that concise by default. I would think that could "break" a lot of things for developers who didn't know to make prompt changes after upgrading to 5.6. What if you were expecting GPT to be as wordy as it usually is? Then suddenly your output is not wordy enough?
Smells like OpenAI trying its best to stave off financial armageddon for another few months. Then again, I'm not sure why they chose to waste so much output computation on verbal diarrhea all this time up to now.
If you conceptualize this as “there is an appropriate amount of brevity for each situation” then it would be expected for a better model to use different amounts of brevity if it gets better at determining the appropriate amount.
My view is that popular models by default output wildly excessive amounts of prose for nearly every use case, so if this changes in a new model that’s a pure win.
The models don't get better, except when a new one is released. Their performance depends solely on the model training before release and how well you curate the context you feed it. That's it. Contrary to popular belief these things are not intelligent.
It seems like the way brevity instructions have changed is mis-aligned with how most people would expect to use them or are currently using them.
Here's the example they give:
> Instead of asking for the shortest possible answer, replace brevity instructions with prioritization:
> Lead with the conclusion. Include the evidence needed to support it, any material
caveat, and the next action. Omit secondary detail and repetition.
> Keep all required facts, decisions, caveats, and next steps. Trim introductions,
repetition, generic reassurance, and optional background first.
Generally speaking, when I ask for a short answer, I want a short answer because I'm not really willing to read through a bunch of bullshit to get to a summary. Putting the onus back on me to assume what the model will return and write a longer prompt detailing exactly what information I want completely misses the point of why I'm asking for a short answer in the first place.
I would presume (perhaps falsely?) that an instruction like this would lead to the model presenting a conclusion not supported by the evidence, and potentially backtracking as it then tries to justify said conclusion.
Yes, if deliberation happens, the model should figure out what it wants to say during that phase; but if you're using auto mode, the model is not going to be doing any deliberating half the time. In those cases, the output blathering is the model's only chance for deliberation. It "thinks as it talks", per se.
Given that, I would advise a different approach: let it blather, but then get it to write you a conclusion at the end that the model can guarantee will obviate the need to read any of the blathering.
I.e. advise the model to add an "executive summary" to the end of any non-trivial-in-length response. With some wording to carefully navigate the model between "the summary is itself too long" vs "the summary acts more like clickbait, leaving out necessary detail such that it requires actually reading the blather."
Not sure exactly what that wording would look like. I imagine something like "write your postscript executive summary as if you were a senior CIA intelligence analyst summarizing ground-level reports into a daily digest for the Joint Chiefs of Staff. Take up as little of their time as possible, but ensure that any detail critical to decision-making is retained." (But that phrasing might only be useful if the model is delivering a certain type of response, and actively counter-productive otherwise. This kind of thing is delicate.)
You are absolutely correct. The second suboptimal part of the prompt is this:
> Trim introductions, repetition, generic reassurance, and optional background first.
It's not possible for the model to "trim" those before they've been output, so this is akin to telling it "not think of an elephant or even take the existence of elephants into consideration while solving this problem".
This was a big concern for earlier models, but with modern CoT trained models they should be able to come to the conclusion entirely in the thinking trace.
Oh the number of time LLM will, for example, be giving me the list of bugs it found in code, when I ask it for a review, just to decide there’s no big half way through explaining it.
Yes this is an extremely well known result for exactly the reason you guessed. It's not just abcktracking, asking an LLM to present a conclusion and then justify is also an excellent way to provoke hallucination as the model con concts "any justification that plausibly justifies the words it's already said".
This is the actual reason why openai _invented_ reasoning models, to give them time/space to work out a solution, rather than having to magic a correct solution out of thin air from token 1.
It's less important now that all models do reasoning, but it's still almost always better to make the output come out last rather than first.
It would (and does), yes; but this takes a lot more output tokens than asking for a summary would. The summary approach is only helpful insofar as it can be cheaper than using the thinking model. (You're basically tricking the instant model into thinking, which it can do, after a fashion.)
But, unless your desired output is literally a document for others to read, at the point where you're having a model generate a full, lengthy output multiple times over with revisions, you may as well just turn off auto mode and have it always deliberate (i.e. choose the thinking model explicitly from the model selector.) Then it'll be as messy as it needs to be while deliberating, but give you exactly what you want as output.
(And if your desired output is literally a document for others to read, that you want to interactively draft and polish, then (in the case of ChatGPT specifically) you should not only be explicitly forcing the "thinking" model, but also should be asking it to activate the "canvas" feature from the start. My understanding is that revising a canvas document involves the model emitting something like editing gestures, rather than simply re-streaming the updated chunks of text. This saves a lot of output tokens on large documents.)
The "auto" mode is (AFAICT) a per-conversation-turn router. (Presumably via a preliminary pass through a very fast tiny model that spits out an number for how challenging it thinks the next response might be to compute.)
On high-challenge turns, the auto mode routes to the "thinking" model. But on low-challenge turns, it routes to the "instant" model.
And the "instant" model, by design, has no capacity for deliberation. (If it did, it couldn't guarantee that its responses would begin streaming "instantly.")
The model will still have read the entirety of the document before composing its response. And I believe that even in auto mode, there are thinking tokens behind the scenes.
Replace 2 word instruction ('be concise') with a 38 word instruction.
Human can no longer be concise when asking for a few sentences instead of 20 paragraphs of BS they don't want to read when all they want is a summary to verify the general direction of the prompt-work before digging into the details.
> At least before it would listen to instructions like this.
Would it actually follow them? IME LLMs are incapable of estimating the length of their own output, the total length of the current context, etc. They just make stuff up unless they have external tools that can inspect those things for them.
It sure is suspicious that both Anthropic (adaptive thinking) and OpenAI (Avoid generic brevity instructions) both seem to be suggesting that the best way to improve outcomes is to entirely leave it to them to decide how many tokens get used.
I mean, it's true that it would be ideal of this stuff did just get figured out optimally behind the API, but there is definitely an incentive on their side to burn more tokens.
I wonder if it will do any better than past versions when one begs and pleads for it to get a job done using a concise, modest amount of code (as an expert human developer might), rather than responding to all prompts by shoveling in a large amount of code.
It creates the context of the request without including language or terms that activate additional areas of knowledge not necessary for an accurate reply.
I'm impressed. It feels like a faster Fable (probably due to the more efficient token usage). It performs roughly the same job, just with 4x less steps (gamedev).
Remains to be seen how the "shorter prompts" advice translates to homogeneity/collapse though.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
I used to go to a barber and if you said "cut it short", he cut it really short.
> GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic. Instead of generic instructions such as “Be friendly and warm,” use concrete guidance:
> Be direct and tactful. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
Soo basically, my new 5.6 custom instructions: Be Jeeves and eliminate all friction from my life through immense processing power. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
This will totally make it brain damaged over a certain tasks. Sort of like the same brain damage that prompted OpenAI project managers to destroy ChatGPT.app today.
> can better infer the user’s underlying goal and intended level of work
This is a trap.
It's the optimistic fallacy that poisons all "consumer scale" machine learning products and what's going to effectively ruin these models as they keep chasing it in the same way that web queries were ruined, social media feeds were ruined, and media recommenders were ruined.
For the vendor, optimizing metrics across their whole user base, they always see positive technological progress as their system gets better at making assumptions and accumulating user engagement scores in aggregate. But for the individual user, most of which has some weird tail intent/interest and some of whom have many weird tail intent/interests, the experience quietly but catastrophically degrades. Output/results become more generic, more divergent with the underspecified "weird tail" intent, and more stubbornly hard to ever wrangle towards that "weird tail" altogether.
We've been watching this cycle happen for 20 years now and it's proving hard for anybody to escape because it works so well for the trillion dollar company driving it forward. But while each step might feel ergonomic and welcome to individual users, there's a frog boiling enshitification at play.
In pursuit of output quality and capability (rather than simply the vendor's user count), what we need rather than "makes better guesses" is "presses for more clarity", even where it feels kind of annoying.
Even among human professionals, one of the first hurdles of breaking out of junior tier work is gaining the confidence to press your colleagues and clients to be more specific in their thoughts and expressions despite their desire to have you do it all for them. But they're often coming to you with incomplete, muddy, and conflicting ideas for which there is no safe and correct assumption that you might just run with, and it's your expertise (i.e. relevant "intelligence") that's critical to bringing attention to that. To achieve professional progression, you need to learn to do that and to not just optimize appeasing the ambiguous client/colleague today in exchange for mutual expense tomorrow. To avoid enshitification, which is probably not possible, we need these models to be learning that too.
I agree to an extent but it needs to be balanced. Receiving a half-baked, extremely verbose recap of thinking on benign details with Opus 4.8 or GPT 5.5 feels like an extraordinary loss of quality of experience compared with fable 5.
Yes it shares less, but I think the trade-off is you pay less in tokens and hopefully it's truly just not needing to say things because it truly does just better get what you're saying, think to read X markdown file or GH issue which contains the info, etc.
As long as I can still push back and get it to share its thinking on demand and I'm confident the model isn't actually basing things on poor premises, this is okay for me. I am more productive when not inundated with time-wasting check-ins.
That said, I absolutely lament the loss of the ability to access the thinking - I would happily read the "DANGER DANGER DANGER" internal gremlin thoughts fable 5 makes to verify something if they were accessed, and prefer that to a recap presented only for my benefit.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
RIP Caveman skill. Six month good. Now skill dead.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
When has this ever not been the case? I don't think this is a GPT 5.6 specialty!
Information density of the prompt is the most important factor in my experience.
And interestingly, LLMs seem particularly bad at writing prompts for other LLMs for this reason (you can guide them to be more dense, just speaking by default).
Conciseness is usually a byproduct of information density though.
There was a fad a while back of building insanely long prompts - tens of thousands of tokens - including having models write prompts for themselves. I always thought it was counterproductive, especially if you're going to use the prompt more than a couple of times. (That said, the e.g. Claude Code system prompt is insanely long, so if you genuinely have a lot of information to provide maybe it's beneficial. Like, shorter is better, but you don't want to be under-specified.)
For Gemini 2.5 and ~GPT5.0-5.1, longer prompts with lots of explicit instructions and examples produced better conformance. Seems like heavily second guessing the models started to get counter productive around the end of last year.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
A shorter prompt results in half as much tokens spend? I find this very hard to believe.
If it's anywhere close to the same universe as smaller models in its behavior, a lot of time in "thinking" mode is spent on reiterating on any constraints given in a prompt. So the more constraints you give it, the more tokens it will spend going "Hold on, the prompt said I have to dot my i's and cross my t's. Let me go through my work to check that all the i's are dotted."
> A shorter prompt results in half as much tokens spend? I find this very hard to believe.
Should be relatively easy to test. And if it's true, just first use a very cheap near-SOTA model to first rewrite the prompt to a similar but shorter prompt before sending it to GPT-5.6.
pi.dev for example can control other harnesses.
An example: the other day for example I didn't understand why Claude Code CLI (which I hadn't used in a while) wouldn't let me cut/paste anymore (turns out they apparently fixed some long-standing scrolling and blinking SNAFU, but this modified how mouse selection/paste worked under Xorg but I didn't immediately realized they changed this)... I had to copy/paste the oauth challenge/response for I was logged out (maybe because I hadn't used Claude Code CLI in a while, dunno). But my usual copy/paste wasn't working and I didn't know how to fix it at first. And because I wasn't logged in, I couldn't use Claude Code itself for this.
My prompt was something like: "Screenshot the Claude Code TUI, transform the URL into a link, open that link in a broswer to get the oauth token, copy it character by character by simulating keypresses in the Claude Code CLI".
(remember: I had no idea how to paste with the mouse not with the keyboard, no I know but I was pissed off and wanted to be logged in immediately... So: another model / harness to the rescue).
(for the curious: it decided to use xdotool and use a 50 ms wait between simulated keypresses to copy the oauth token)
This worked just fine. And I that with a cheap model.
I think that just like Linux and Git owned many proprietary software, we'll soon have fully open-source harnesses orchestrating everything and delegating the work to proprietary tools (like "ChatGPT now Codex and vice-versa" and Claude Code)... If proprietary tools are even still needed at all.
Honestly I begin to wonder if they're even needed at all: the models, sure, while waiting for the open-weight ones to beat them. But those proprietary tools trying to lock people in?
I feel like the open source harnesses are already more powerful.
Seeing the dramatic differences in scores just going from high to xhigh is just another demonstration of the bitter lesson: Just keep scaling search and learning.
We are probably going to need a lot more GPUs.
GPT-6 is supposed to be using a much larger base model that just finished pretraining so the "dump another chunk of all written language" approach is still going strong.
The scaling with reasoning models is more and more with things like verifiable rewards (coding and math), in line with bitter lesson and also Sutton invented lots of modern RL.
While I think this is true, remember as we get more efficient we just decide to scale even bigger. So more GPUs, and more efficient.
I agree with the sibling comment, effiency is probably the more important component at this point. We are hitting not just a practical engineering roadblock for scaling with current technology, I think we have definitely hit a financial and logistical roadblock for up scaling with the number of GPUs (on an immediate basis)
Not always, in some cases, changing to a higher reasoning makes the AI doubt itself too much, and skip over the correct answer by overcomplicating the problem and polluting the context.
It would be nice to see on which categories of problems the extra thinking makes it better and on which it makes it worse.
Kind of refreshing though that the "throw more processing at it" scaling we saw in the 90s has returned in a different way. For a while we were really bottlenecked in our advances by relatively low levels of parallelism (most software used by your average user doesn't scale cleanly with more than a few threads).
I mean, theoretically you can solve every finitary problem with a brute force solution...
Richard Sutton specifically states that the search has to be smart. We know that the brain uses recurrent connections and is shallow. I think a lot more money has to go into architecture. Feed Forward transformers can only scale so far
This isn’t really how it works anymore. Agents rely heavily on tool use and the agentic harness to perform tasks. Pre-training is no longer very effective.
Moreover we've known for quite a while now that glial cells also participate in cognition and moderate learning (e.g.: [1]). When you take those connections into account the numbers get really staggering. 85 billion glial cells with trillions of protein channels facilitating communication between the glial syncytium [2].
The real question is not how many "weights" the human brain has (neurons+synapses may or may not translate into "weights", and brain might be also inefficient for what it is), but rather how much evolutionary and social "compute" was necessary to pack everything into that capacity.
For intelligence, I expect the next breakthrough to be colocation of memory and compute in the same chip. And we'll need much more of this memory, probably a few petabytes.
I'm surprised it is that low. Are not all top AI labs "cheating" and workaround LLMs's low sample efficiency by hiring people to generate more data points - similar problems with answers, so they can train models on those and improve scores? A good benchmark for general intelligence probably should be a complete black box, no sample data given/leaked at all.
Very interesting. My prediction is that Mythos would outperform Sol.
Also what does this tell about Yann LeCuns whole world model theory? Bro has been going on and on about it. He has made multiple wrong predictions on the trajectory of LLMs.
At some point his claim should be fully falsified no?
Are you joking? They spend billions of dollars training LLMs to get a 7.8% on arc agi 3 whereas DINO models are near sota in image classification, provide meaningful embeddings to the point where image segmentation is just PCA. The spend on DINO cannot be more than five million (correct me if I'm wrong)
His main anti-LLM predictions have been consistently either wrong or misleading.
There's many ways to skin a cat so you can probably do something with a JEPA approach as well, but I doubt he actually catches up to having agents on the level of where Anthropic/OpenAI will be at any point.
His main LLM predictions have almost nothing to do with Arc AGI...
What exactly was he dead wrong about that is proven by any of this?
GPT getting better has absolutely nothing to do with completely disproving anything LeCun has been saying.
He never said LLMs couldn't get better. He never said they couldn't score 7.6% on Arc AGI 3.
He's merely said they don't think, and you probably want something that actually thinks if you want a model that can be trained cheaply on a small amount of data and provide a ton of value.
Spending $5B to train a model that scores better than an older model does not disprove any of that in any way.
He said years ago even 'GPT 5000' couldnt do things that they ended up doing fine a month later, let alone by 5000. His later predictions are just moving that goal post including towards them not being able to do more general, harder problems of which Arc AGI is a counter-example.
That does not at all look cherry picked or taken out of context...
LeCun's ideas cannot be reduced to a 6 second clip...
You're missing the forrest for the trees, taking a singular example of a problem and thinking that if an LLM can solve the singular example it completely disproves LeCun is comical...
You can watch the whole Lex Friedman interview, it's on youtube. It's not out of context at all. He goes on about how LLMs will never be able to do things that they do trivially. And he has just doubled down for years.
Ive read and watched more of his interviews and lectures it seems, it feels like you just have a rosier idea of his views than the views he repeatedly presents.
I think you misunderstand what he thinks LLMs can and can't do. He says repeatedly that LLMs may be great for generating text and code, but that a fundamental model of artificial intelligence should be able to perceive and use information outside of text. Also, that a supervised learning approach is not exactly representative for how we learn. That it's a small piece but we largely learn with unsupervised learning. His main criticisms of LLMs are that they are supervised, probabilistic, and learn largely from text instead of observations. His claims about performance come downstream and I'd still argue that he's been (somewhat) right about those as well.
That LLMs don't have common sense and don't have good physical reasoning abilities; that you can't scale LLMs all the way to AGI; or that they can't predict the consequences of their actions which is the foundation of agentic behavior all seem like still (mostly) accurate predictions to me.
While LeCun has his share of problems, I think largely his criticisms on LLMs are more right than wrong. What remains to see is how good JEPA can be at filling in the gaps left behind from the brittleness of LLMS.
To me LLMs have gotten better since 2024, but their fundamental flaws still seem there.
They hallucinate when it comes to really challenging tasks such as math proofs. They still do not reuse code well and will rewrite functions instead of perusing the standard library.
But this is good news. LLMs are awesome and they are only the first step towards AI being applied everywhere. They are a Model T
> What exactly was he dead wrong about that is proven by any of this?
He said as you need more and more tokens models will fall apart because each additional token is a chance for a mistake and they will just exponentially fall apart. But in practice models have learned to identify and self-correct mistakes and if you look at the graphs more inference reasoning tokens almost always give far better accuracy.
That's true, but he's still correct, it's just that the context is now so large that only people using agent loops see "context rot"
His other criticism of LLMs that I like better is that they try to predict tokens instead of learned embeddings. Tokens are arbitrary and in order to decode LLMs you need technical analysis (see mechanistic interpretability).
With JEPA models so far, it seems that PCA on latent vectors suffices.
tldr: embeddings have a lot more room for improvement
From CoreWeave, at current prices (~$2.46/hr spot to ~$6.16/hr on demand) would correspond to $22M–$55M.
The dataset is really where the cost is though - they used LVD-1689M - 1.6B images of curated web data from roughly 17B instagram images. This probably cost a huge amount of hours in human annotation, compute for algorithmic filtering, etc and not to mention probably a 20-50 person team working on this model.
You might want to change assumptions about how expensive these models are.
Ok long time Claude Code user here; lately I've started to realize there's other great models out there I should be trying, but I'm hesitant to leave Claude Code behind for something new.
What's the consensus today on codex vs claude code, does it really matter anymore?
Codex has arguably been better than Claude Code for months now, but it's flown under the radar because it just didn't capture the same viral marketing effect and OpenAI in general has had more optics / PR issues than Anthropic amongst the online developer crowd. I use the word "better" not in the sense that the underlying GPT models are fundamentally smarter or more intelligent, but rather that as a product Codex is just simpler, cheaper, and abundantly reliable and low-drama.
I’d argue the opposite. I’ve switched back and forth from one to the other and Opus/Fable has been constantly better than any GPT in my daily work. It’s a bit slower but it does the things right, with as little code as possible, some comments where needed. Codex is faster but you always have to correct it because it got something wrong; it writes tons of code ("let me add a small helper") with obvious comments.
Purely anecdotally the one persistent issue I have with LLMs writing code is that they are absolutely paranoid and add a load of indirection and defensive crap and even if you prompt to avoid that it will often require manual steering to remove the cruft.
recent gpts are horrendous for this, whereas recent claudes have a tic where they incessantly add useless comments referring to previous changes and will use multiple single-line comments instead of a standard multi-line docblock.
Sounds like you are talking past each other. GP is saying the harness of codex is higher quality, which I can believe, even if the models are not as good as Opus/Fable.
i don't think so, i think it's 50% what work people are doing, 50% vibes. my experience with 5.5 is i like it more and get better results than 4.8/fable. which isn't to say i think it's a strictly better model, just been working better for me.
I really love the Opus/Fable models but I'm honestly sick to death of the buggy product. The CLI always has some weird issue. Right now it doesn't even output messages before tool calls, it just swallows them and they disappear.
I don't like OpenAI as a company, but they appear to have QA, and that is probably enough to get me to switch.
There was an issue on Claude Code the other day where it would only wait 60 seconds when it had asked a set of questions, then if it didn't get a response from the user it would just continue however it thought was best. Completely unusable. It took them nearly 48 hours to merge a fix.
I have both as well. I trust the output of Claude to a higher degree than what I get with Codex. I always have claude review codex output. That being said, I find gpt 5.5 more generically useful at a wider breadth of tasks. Straight coding though, it's no contest.
Obligatory YMMV, maybe your prompting style fits gpt better. We forget that this matters a lot
> Codex is faster but you always have to correct it because it got something wrong
this has been my experience with Codex as well, and I have to fix its mistakes every single time. But recently, I literally threw away three hours of work because it kept adding hundreds of lines to my code base. When I restarted the entire work using Fable and Opus, it was like night and day.
Agreed. GPT 5.5 will come up with more straightforward solutions with far fewer tokens than Claude. Also, the usage limits are much more generous for Codex than Claude Code for the same monthly plan.
That's a strange statement... It's been true for a while now that OpenAI has had much more generous limits than Anthropic on their subscription plans. And with the Fable ban/guardrails disaster, there has been a lot of frustration from people in these comment sections. And Anthropic fucked up Claude Code pretty badly for a couple of weeks during the 4.6/4.7/4.8 transition, which again was widely publicized. And they got a lot of flack over not allowing other harnesses anymore. And ChatGPT got some pretty viral wins on model intelligence when they cracked the high profile Erdos problem.
If anything the online optics have been bad for Anthropic for the last half year. OpenAI doesn't have optics issues, from my point of view they simply have the issue that they are the least trustworthy player at the frontier. The way they pivoted from their original mission is truly breathtaking, especially coming in gloatingly to take the government contract when Anthropic got kicked out for insisting the government does not use their systems for mass surveillance or autonomous weapons systems. You understand what that means, right? OpenAI models are now actively used/developed for mass surveilance and/or autonomous weapons systems.
I know there are plenty here who seem to value their own ability to use these models cheaply above all other considerations. Then OpenAI is a great choice, and much less restrictive than Anthropic. But their problem is not on the optics. It's on the substance.
I really want a good Claude Design competitor in Codex, it's hard to use the others after getting used to it and yet I find anthropic's model to have a much worse understanding of what looks good or not than OpenAI or Google models.
Nudged by this thread, I've decided to switch from Claude to Codex for a bit to see what happens. But...I immediately became lost in their marketing vortex of confusion on plans and pricing. Anyone care to tell me which plan I should be using? On the other side I use the $100 Claude Code plan. We actually have a "Business" ChatGPT subscription already, which seems to be $50/mo/seat. OpenAI's web site offers a set of individual subscriptions (for parity with CC presumably) which I suspect weren't available when we signed up for ChatGPT. I think that in turn happened due to some web site feature it didn't allow for free users (uploading PDFs, something like that). Perhaps I should switch from that business account to an individual subscription for Codex?
Test-drive it with an individual Pro account (5x or 20x) for a month. Download the Codex CLI client from https://github.com/openai/codex and auth it in the browser via the URL it provides. Set the model to 5.6-Sol and effort to max.
Honestly it’s the usage limits that are so generous that makes codex worth it even if it may not be exactly as powerful as Claude. The peace of mind that you can try a lot of things and make huge refactors and run extensive redundant tests without running out of tokens just makes the whole thing a much better experience. I tried coding with Deepseek and it was pretty terrible so the only reason codex works is because its abilities are close to or on par with Claude.
I use both. Both are great. But in terms of Desktop Apps I think Codex has the better UI. It's more straightforward, just works, and has small conveniences like the open in editor icon.
Claude's very bloated and convoluted by comparison. Maybe you need the bloat (Claude Design), but I prefer the more razor's edge efficiency of Codex.
Model wise, I can't really tell. They all do what I want them to do most of the time and go off the rails occasionally. The question is increasingly becoming who's faster and cheaper and gives me more tokens, not who's better.
There is so much less drama involved with the Codex world. You don't realize how oppressive CC is until you've escaped it. Outages, weird restrictions, degradation, accelerated usage, etc etc etc.
Totally. My experience as well. After some time with codex you're like come on Claude can you just stfu! Haha. I now almost always instruct Claude with specific length requirements when I ask questions. Otherwise, it just blathers and blathers in the most annoying of ways. "Oppressive" is spot on in my opinion
I'll agree and expand on "weird restrictions" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc
Let alone getting banned out right with no reason, zero updates after weeks, and not even being able to download your chat history (despite the feature being available (I assume they vibe coded it and it does not work!). My story below;
Um, the 'codex world' is the OpenAI world and there is a ton of drama and product confusion there!
Anthropic has certainly had some drama inflicted on them by the US administration, but otherwise they have just had heads down and executed with great focus. That is why they have succeeded.
I've been using Claude Code, Codex, Gemini (now Antigravity) at the same time for half year now, ever since I dipped my toe into agentic coding. I'd say in general Claude Code and Codex are equally powerful, Gemini is lagging behind.
One thing I appreciate with Codex is, OpenAI nowadays sometimes just gives you quota resets you can bank, so when you use up weekly quota before the week ends, you could just reset the quota, to continue using Codex. I've been much less anxious about Codex quota because of this perk. I just used one reset in the bank yesterday, and still have 3 resets left. Whereas with Claude, when you've used 95% quota 3 days before the week ends, you'd be much more anxious.
On the other hand, Claude Code's /remote-control mechanism is extremely helpful when I am running it in the cloud and wants to monitor it or control it on my phone. Codex currently doesn't support this kind of usage. Codex only allows you to use your phone to connect to a session on your desktop, not in the cloud.
Yes - Anthropic badly needs this same "here's a reset, use it when you want".
It's vastly better this way. Sure, it may impact the bottom line but it's a huge customer satisfaction win.
When Anthropic randomly resets me and I've only used 2%, that's worthless. When OpenAI tells me I have 3 resets available to use whenever I want - it's wonderful.
I have used claude and codex extensively but only from their CLI app (heavily sandboxed using rootless podman, network filtering, etc), so I don't really know what I'm missing with the GUI apps.
One killer feature that Claude has, and AFAIK Codex still lacks, is the ability to start a session in the terminal and then hand it off (actually just remotely control it), from the iOS app.
Last time I tried Codex on iOS it required a ton of set up to link a github project etc. The way claude lets me remote into a session I've already started on my actual machine is much better IMHO.
They’ve addressed that. Codex in the ChatGPT app on iOS is way better than Claude Code now.
You sign in the Codex app on your Mac same on iOS and are able to completely control your sessions - fork, side chats, plugins - everything.
It’s really great i often work through it. And you can connect any number of Codex instances on any number of macs and then manage them all through the iOS app.
maybe I'm misunderstanding but I don't want to sign into the app on a mac - I want to run the CLI on a headless linux server and control the sessions from my iPhone. Does Codex allow that now?
1. Run `codex remote-control --help` directly on your Linux server.
2. From the desktop app, connect to your Linux box, start Codex there, and make it remotely controllable.
Hmm, I don't have a desktop computer. I prefer my laptop be used for other purposes, and can sleep when not in use, instead of running a coding agent 24x7. That's why I prefer running coding agents in the cloud.
You can do that too. Start working locally, then just do /handoff to transfer your session to the cloud and then work through the Codex app on your phone.
One thing I noticed you can't use codex installed via npm, but it will tell you that.
Ymmv I'm a pretty simple user and do things in small manageable chunks. I try keep context < 10% before I fire off a task so I don't use many skills etc.
I recommend trying Codex too. In fact, I recommend running them side-by-side if you have the budget, e.g. have both independently plan the same feature or implement in a different worktree, or have them critique each other's work.
I personally find GPT-5.5 to be a better programmer than Opus 4.8, it is extremely thorough, but I don't like the code it generates ("austere"), and find Opus 4.8 to write more "human friendly" code. The programming comments GPT-5.5 makes is pretty awful where-as Opus 4.8 is good. I feel like Opus 4.8 is better at grasping my intention than GPT-5.5, and honestly find GPT-5.5 to be kind of "autistic". I do prefer the language (not the writing) of GPT-5.5, as I find the philosophical flowery language of Opus 4.8 kind of annoying.
I have only managed to try Fable 5 a little bit, which feels like a much more generally smarter version of Opus 4.8, that is much better a programming and grasping your intention, and I think even the intention of your code, and is _really_ good at spotting bugs or problems with logic in your code. It feels wicked smart but is extemely expensive. It feels smart in the sense like it has a "bigger brain" and is much more sensitive to subtleties/details.
These are different "brains", have different "personalities", etc. I think the best thing is to develop a feeling for it yourself.
I haven't tried Codex yet, but I for my tasks GPT-5.5 may correctly point to a proper direction but its code feels a bit weird. Opus 4.8 is way better in coding, and actually it's the only one who could catch very very sophisticated bug in a large codebase (I tried different models including GPT-5.5 and DeepSeek). Interestingly Gemma 4 under opencode running locally performs not bad at all, it's far yet from DeepSeek level, but it manages to understand tools quite well, and code quality is pretty good. So, for simple coding projects I can say local models already won. It's amazing how smart open models of desktop size have become today. I mean it's quite plausible to manage small codebase today relying on only open tools and local models, you don't need any subscription to produce high quality code, but yes I assume you already experienced and know what you're doing :)
Claude Code fan here... Codex is very good. Sometimes better. The killer feature is price.
After 6+ months of exclusive Claude Code usage, I was begrudgingly forced to try Codex once Anthropic rejiggered their limits such that I kept maxing out my $200/mo plan in just a few days. These days I pay both $200/mo plans, and it's just about enough to get me through a week's work (small game studio - infinite code to write!)
In terms of ability to ship? Easily tenfold. We literally ship 10 times more than before AI. This does not, however, translate into a tenfold increase in actual business success, of course :)
Genuine question/not a critique-are you actually reviewing all that code or just sending it and hoping for the best? I just can't imagine someone is reading/reviewing that much code every day, but maybe I'm wrong?
Like before AI, the scrutiny varies with the sensitivity of the area being edited.
Simple UI change? I do an AI review, but otherwise neither read nor write the code. The models are good enough they write better UI code than me, 9 out of 10 times. Not always the more idiomatic, but usually safer and more correct.
Change to our core data plane? I might spend 2-3 times more effort reviewing it than before AI. Yes, I go more slowly than pre-AI. Many more reviews, many more angles considered, including both human and (lots of) AI review cycles.
Most code is not that critical, and AI is also scarily good at writing tests. We also spend considerably more time paying down tech debt and testing thanks to AI, now that the cost is near-zero.
Net: I spend 10-25X less time on low-risk changes. I often direct (or at least approve) the implementation approach, but I rarely read this code. I spend 2-3X more time on high-risk changes. In both cases, I never write code "by hand". Since about November, I've had no reason to actually edit code in a code editor (perhaps maybe except .env files, which we don't allow agents to edit for obvious reasons).
AI is a tool. You can use it to go fast recklessly, or you can use it to go slow with confidence. Just like before AI... the skill and art of engineering is knowing when to do which.
Now we have various Opus+ level models (Opus/Fable, Grok 4.5, GPT 5.6) I prefer to focus on price/speed and harness as models are all generally good enough for coding. (Fable is overkill for 90% of work but is still level above). So I use Grok Build with 4.5 as its VERY fast and cheap, Codex is next best for me with sol/lunar 5.6. and Claude Code Fable for the 10% of tasks that need that level of reasoning. However I find Claude Code harness responsiveness much less than other two (all TUI versions) I wish they would fix this.
It never really mattered (except when codex was very new). If anything, codex's remote session integration is better, so outside of some "ultracode" orchestration bells/whistles where Claude Code is ahead, I think Codex is a better tool.
Codex has been good for a long time, more expensive but very focused on efficiency. Working with it feels faster and more to the point than Opus models and I trust it more with long-running jobs. Also regular resets vs being at the whim of Anthropic drama all the time is hella nice.
They've discovered it's a good marketing strategy. Whenever there's an outage, or a new launch, there's often a reset with it, which helps keep people engaged with OAI / Tibo and reduces churn.
They've also introduced banked resets, which are really clever. If you have a $200/month plan and three banked resets, you're not churning because you will overweight giving up those resets (loss aversion theory).
They're different models with different philosophies behind them. This is anecdotal with a user group of 1, but in my experience:
Claude has a stronger personality and is more creative. If you give it vague instructions, it's better at filling in the blanks with reasonable ideas.
GPT-5.5 is better at following instructions. If you know exactly what you want, it will do it without going off the rails. It's also less likely to imply that you're dumb, but I don't really care about that. Some people do.
I’ve found that Claude is very literal. When I talk to 5.5 it gets what i want it to do, when I talk to Opus 4.8 it does what I say literally and doesn’t get the intent behind it.
Personally, I started using openai models to mess with other harnesses. I was pretty oppositional to CC and how they don't let you kinda plug and play freely, or give transparency into -p usage with other harnesses. So i mix and match a bunch of openai and some chinese models im trying out into opencode. I keep hearing codex is great, on the tier of current CC, I've tried it and it just ate my entire 5 hour usage window looping without asking for clarification on something and none of it was usable. that was the only time i tried codex as i could got that same task done with maybe 20% of my window with my existing openai opencode workflow.
I had put a decent amount of effort into setting up that initial codex attempt and it went so poorly that i've been entirely uninterested in trying again. This was maybe a month or so ago, and i know stuff moves fast, but for me, i like the models, dont care for the harness.
I personally use opencode so I can swap between models and try different options. I'd say I prefer claude (fable and opus 4.8) so far, but curious to see where gpt 5.6 lands.
For personal stuff, I've been pretty happy with chatgpt's $20 plan. I believe it has considerably higher limits than claude's $20 plan, and it's enough for the personal stuff I play with (hermes, and some small coding stuff). Also allows me to keep up to date on openai models.
I've been using it with hermes and some coding (with opencode), and I am getting a LOT more than one feature out of it, but the work is spread throughout the week.
I can't tell the difference between Fable and GPT 5.5. I tried Fable while it was in trial $20 mode, used up my whole quota, and it was great, but as soon as I went back to GPT 5.5, everything was the same.
But what I love about Openai is that they still let you hook OTHER harnesses up to a subscription. My Pi setup has been built up for a few months now into exactly what I want and moving over to CC or even Codex is really annoying.
Caveat: I vibe code in tiny little chunks. I see what I want to do, and exactly how I want it done, then prompt that, refine, what was output, then repeat. I bet Fable is better at building a whole app from a 2-sentence prompt; but that's just not important to me at all.
Don't know about consensus, but I personally still find Opus to be better for sniffing codebase intent and checking things as a whole, while Codex seems more detail-oriented for individual files.
Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried.
Claude lost my trust around February this year when the plan would say nonsensical things as "delete this method" that was clearly a key method on that part of the codebase.
For personal projects I am using Codex 20$ plan and when that is over I use DeepSeek which is insanely good for the cost.
You can however for now use wrappers which are not harnesses such as T3Code though. They were going to cut under the Programmatic API, but have at least temporarily walked it back.
You absolutely can; they are not banning anymore. The bigger problem is that subscription versions of the models are way crappier than when the "same" model is hit via API (Bedrock/Vertex)
You can also make it not count against extra usage.
OpenCode docs show it because Anthropic specifically ambushed them with a PR to remove support so simpletons can't use it easily.
How does that work? Doesn't that mean Microsoft/Amazon/Google have full de facto access to OAI's and Anthropic's model weights and operational processes?
This is one reason it surprised me that Anthropic decided to run stuff on Musk's hardware. It seems overwhelmingly likely that the new Grok release is informed by what Musk has been able to learn from that relationship.
They aren't banning it anymore, they just make it count as "extra usage". e.g. you're paying for every token in addition to your subscription.
Further, the claim that the subscription "version" of the model is worse sounds like bullshit (and the sort of anecdotal nonsense that you see on sites like this). Do you have anything substantiating this?
> What's the consensus today on codex vs claude code, does it really matter anymore?
Consensus is probably the wrong word for the popular opinions reflected in HN that you might get.
I would recommend that you have 2 of each at all times when it comes to AI so you don't necessarily become overly locked to quirks of one thing. You'll soon realize that things move so fast that you just start internalizing common patterns instead of depending on one specific vendor.
I recommend that you try pi and codex besides claude, to get your own feel for it.
I spent the last couple days switching because Anthropic keeps locking stuff behind API pricing. OpenAI lets you do anything with your sub right now. I'm building headless and web interfaces around Pi.dev. I had this previously with Claude Code but they are going to lock away all those features. I think the Claude does a better job at being proactive to solving things, but I'm going to keep tweaking my harness to nudge gpt to do more in it's turn. Not sure!
Claude Code is not the model, it's the harness. You can use any model you want with Claude Code to varying degrees of success. I use Qwen3.6-27b daily with Claude Code as an example.
Set yourself up to be able to try / switch between models easily. I was a claude only user and just have my user level AGENTS.md for codex and others simply point at my user CLAUDE.md. Have a script that syncs my skills (just directories) between all models. Also, if you want to use /simplify or similar from claude in another model, you can ask claude for the prompt and put that in a skill for the other models.
People work on different things, fail to mention their field of usage in the comments, and then misunderstand the experiences of others who do the same. Repeat ad nauseam.
More literal, less fluid verbally, harder time understanding nuance, more correct code, fewer bugs. Less pretty UI. I switch back and forth but find I have less 'clean up' work with codex; more upfront communication though to properly specify. High hopes for 5.6!
Personally I use Open Code with a copilot sub. Then all models are available in my session with just a /model and /variants command combo. Makes it super low friction to try different models & combos (my favourite right now is DeepSeek V4 Flash for initial PRD then Fable 5 high for implementation).
I consistently have better results with Codex for the work that I do. People have been saying that for six months, but until 5.4 the experience was sufficiently slower that it wasn't worth the switch. Making the switch was frictionless. Give it a try
I had great results combining the two. If you (or your employer) can afford then you can ping-pong the models in the plan phase (not really ping-pong as humans should get a say too) and then let one implement and the other review. I got better results working this way than just to stick to a single model.
I use both. Not because I am cool, but because it is cost effective for personal projects with two $20 / month plans. It is also nice to be able to see what the state of the art is like for both.
Personally, I find it very interchangeable. I open codex --yolo or claude with whatever there yolo flag is (have an alias).
I use Conductor pretty much exclusively and it makes it incredibly easy to try different models, even within the same workspace - definitely recommend giving it a shot. Whenever I'm forced to use the Claude Code app directly it just seems woefully inadequate compared to Conductor
I'm also a long-time Claude Code user here, though the last 3 weeks I've been doing loops having claude use codex to review until they reach consensus; uses tons of tokens but the result is really good.
I'm trying Codex as my primary the last day or so, because I'm at 98% use and reset in 3 days on Claude. I'm worried about a lot of our skills and CLAUDE.mds and the like getting lost unless I migrate them, but otherwise codex seems to be working great.
If you can afford it and you have something to justify the expense, I would get both. they're interesting to run side by side, you can hand things off from one to the other. Pretty neat. Unfortunately now I just want to have both :(
The harness is so much better than cc which is a buggy mess. Gpt is also way faster than Claude. I’ve been using gpt for a while now and I know a lot of people that swapped away from Anthropic for multiple reasons. However - fable still seems to be the best coding agent, it’s just slow and the harness sucks. So I still use it in some rare cases like to review codex. I’m hoping 5.6 lets me drop it entirely.
I have one non technical people in my firm using it. One is using it to assist with editing books, basically using it to gather up manuscripts from e-mail / Google Doc etc. submissions, and then switch models between a cheap one and Opus (for actually analysing the manuscript).
The other non-technical person has done really surprising things with it AI, like a long-running GPT 5.5 Pro chat session which is basically her expense tracker - it has an .xlsx file "carried" in the chat, and she just tells ChatGPT (or scans a receipt) whenever she has a new expense, and then prompts it in natural language when she needs a report. I'm looking forward to seeing what she can do with omp.
I figured pi itself would be the best harness because it's barebones and you make it what you want. omp is to pi what doom is to emacs is what lazyvim is to neovim.
The fact that I thought that this was amp misspelled until i someone validate omp and the checked myself indicates it's a subjective assertion at best.
My experience is that Codex's auto review is extremely costly, with $20 on both sides, I can run CC with auto mode for longer than with Codex's auto review enabled. Also in my own experience Claude's usage is actually bigger than Codex, but I am not sure if that's due to I stick to 5.5 with Codex while keep Sonnet as the default to orchestrate other models in CC.
IME it entirely depends on your work. I find myself using both daily for different things.
Codex with GPT 5.5 is much better at general SWE tasks but Claude Code with Opus is far better at complex reasoning tasks like reading and summarizing research papers, replicating experiments, identifying research gaps and proposing interesting follow ups.
Codex app is a much different experience than CC CLI. I would try it out for a couple days with the new model suite and see what you prefer after that.
I sub both codex and claude at 20x. I like opus+fable more than gpt5.5 because it seems gpt tries to finish tasks by leaving any ambiguity unresolved. claude seems better at surfacing open questions.
This is using the same AGENTS.md prompts, which were designed firstly for Claude use, so maybe it's something that could be optimized better if I understood gpt as well?
If you can afford to test it seriously, running both in parallel, it's worth a test to see which you prefer. If you can't, don't bother. You're not likely missing anything since they are close to personal preference with most people I know who have meaningfully tried both preferring Claude
Last time I tested Codex on a cheap plan, it barely lasted an hour? I think this was for the $20 plan. I was afraid to try the more expensive plan after that. Not sure, I might just outright rip my Claude Code bandaid if the current usage quotas do die off after the 17th or whatever date they said they would "return on".
The answer is it depends. Claude's generally better at frontend and debugging tasks, while Codex is stronger at backend features and exploratory work. They have very different coding styles and thus very different strengths.
I left Claude for Codex months ago. I was an early Claude Code adopter but I have found Codex consistently better since about the February time frame. And far more reliable.
It's more diligent and empirical and results focused, and less creative. It sometimes needs a kick to avoid a Zeno's paradox of incremental steps to get to the goal. But it produces more reliable code with fewer race conditions, unhandled negative cases, etc.
It's also better value from a $$ POV, or at least has been. This fluctuates a bit.
You're also free to use your Codex subscription with other harnesses, like opencode, etc. Unlike Anthropic. Plays better with others.
Not sure there's going to be a consensus, but I can tell you that when i have codex review claude-written code, it finds important gaps and fixes. The reverse is also true. Both are powerful, but even better when used in combination
For me the biggest shift was using Deepseek through an American provider with reasonix as the harness, making cache hits at a rate of practically free.
like others said in the thread: much less drama and i'll add much less attitude from the company and the models, overall i'm having much calmer experience with codex, hope it stays that way
They are both excellent but excel in different areas. Fable is super super proactive and great for doing a LOT of work with a single prompt, also for creative work.
Codex is more details focused, often catches wonky bugs and correctness issues that Fable misses, feels more terse and less "friendly", more like a stern senior engineer versus a friendly talkative engineer (Claude). Codex is also better if you're already an engineer, Claude is better for non-engineers. I.e. Codex works better if you know exactly what you want and know the right way of explaining it.
In my projects, Claude writes and Codex reviews, and I've had a lot of code I've been very happy with out of that, although as of today, Grok _also_ reviews, and finds interesting new stuff.
- the tool calls and results are very legible, I can click them and see the progress
- no one talks about this but the tool call and response notification are handled much more elegantly in Codex. In Claude Code, it is handled in a clunky way using loops which always causes some delay
- you can steer the conversation midway in Codex
- /side is underrated (/btw is the equivalent and is much worse in Claude Code)
- I have to admit subagents are handled better in Claude Code
With the exception of Fable which is going away anyway, Codex is better especially after the last couple Opus releases. It’s also no longer slower than Claude.
You get much more generous usage from the 20x plan.
And you get far better uptime.
If benchmarks and early tester impressions are accurate, you also get access to Fable level capability at greater speed and lower cost (included in subscription).
The fact that they already extended subscription Fable once would suggest it won’t be solely locked behind API next week, but at the same time it really does look like they are doing everything they can to avoid serving it continuously at scale.
Knowing Anthropic, this unfortunately might end up meaning a quietly quantized Fable on subscription.
Can anyone explain this "quietly quantized" model idea to me from a business perspective?
Coca-Cola doesn't "quietly water down" its product to save a few bucks. They know people will take a sip, say "oh that's not what i wanted", and go buy a Pepsi.
If they serve me a quantized Fable, I'm just going to think Fable sucks and go get my tokens elsewhere. What's the point?
> Coca-Cola doesn't "quietly water down" its product to save a few bucks. They know people will take a sip, say "oh that's not what i wanted", and go buy a Pepsi.
Coca-Cola is also mostly measurable and reverse-engineerable.
The Claude models are black boxes, and actively curtail distilling efforts.
Pepsi may not water down its product, but your local diner may well decide to put a little less ice-cream in its milkshakes if it thinks it can get away with it and most people won't be able to tell.
Same here. I find the design, architecture, system design discussion to be better on Claude, but after Opus 4.6 I switched over to Codex for actual coding and love the results. I use both via the CLI and generally tell Claude to output the result of our decisions as a markdown that will be easy to read and implement by an agentic coding tool. Then I fire up Codex and read said markdown as the input of the session and way to build all the appropriate context needed. I see this as a way to step into letting the agents go run on their own and interact with each other, but I still like to steer so I put these manual steps in the flow. Letting the agents go off on their own and one shot big chunks is not reliable enough yet imo.
Pi is so “unbloated” that it’s extra effort to use. You can decide how much work to put into it. I get the trade off. But this is a big jump from CC. I’d recommend some middle ground like opencode.
Even simpler, use Cursor with any frontier model. I see others sweat to add enough context to Claude Code while Cursor has a ton of contextual awareness, uses subagents automatically and is significantly faster with no drop off I have found. I'm not sure why devs are so enamored with living in the CLI, but Cursor has one of those too.
Similar to running arch Linux. Many people do need to. But many people just like tinkering. Tinkering can lead to positive outcomes but it’s usually not “doing work”.
It's worth trying out OpenCode, then oh-my-pi, and also the commercial harnesses like Codex. (I haven't yet bothered to try Antigravity and have no interest in Gemini-cli now that it's not available except on expensive plans.)
pi is also worth tinkering with, particularly if you have an eye towards automating some things.
It honestly baffles me how people can ask a question like this and get such a wide spectrum of answers in response. It's all so much based on vibes and anecdotal evidence. I've not really noticed much of a difference in capability since Opus 4.6 and I've used a ton of different models. They all work pretty damn well for me.
I've subscribed to ChatGPT/Codex for over a year and tried a Claude sub twice 1 month each, with a gap of several months in between.
I tried them both side by side, mostly for reviewing existing Godot/GDScript code, or sometimes generating Swift Mac apps, including converting ancient relics I wrote eons ago in Visual Basic on Windows
Besides the useless "This is good" findings while reviewing and the excessive "oops you're right" backtracking, Claude's atrocious UX and borderline "spyware" make me never want to try an Anthropic product again for a long long while.
It's not clear replies to this thread aren't openAI employees or incentivized influencers, but every benchmark has gpt-5.5 underperforming opus 4.8, sometimes by as much as 10%.
Funny to see that they did not include Fable 5 in their GeneBench and LifeSciBench comparisons because "it does not answer advanced biology questions and refuses the majority of questions in this eval".
It's so absurdly sensitive. It bailed out earlier today working on a TypeScript client for a sensor network API which happens to include some temperature and pH sensors for tanks, which yes, are used for biology experiments. But wow, we're degrees of separation from the actual biology work.
It's making it very hard to justify even trying to use Fable. When it works, awesome; it's legitimately good. But I can't trust it to do a task without deferring to Opus and that's really annoying at times. I want to know what I'm getting up front, not after the fact.
I mean it's a fucking joke, I kept getting refusals on a code base I wasn't familiar with and it was literally just because there are some vars named DNA. Just absolutely stupid.
Anthropic just refuses to allow Fable to properly code review my projects. It's so obnoxious. If OpenAI's Fable equivalent is better at this, that'll get me to cancel my Anthropic subscription and switch.
Well it seems like they removed quite a few 3rd party benchmarks they used for GPT-5.5 release where Opus 4.7 was better and added many new benchmarks created by them where conviniently GPT leads.
Not at all, we love them all with Chinese labs. And wish them to continue competing and not winning. That is how we get best models, lower prices and better availability.
Comparing this to other models, I find it similar to GPT-5.5 and a bit behind Sonnet 5. You can see how other models fared here: https://senko.net/vibecode-bench/ (you can also fetch the prompt and the the 5.6 Terra resulting code on from that page).
I don't have access to Sol yet (on a Plus sub, which should get it according to what I've read), so can't do the more interesting test. I'll update the above page as soon as I get access - hopefully soon.
It even has enemies! (I'm not too mad about it not following my instructions because it can be fun to play :) And I generated that from Claude Code on my phone.
So the measure of a model is how well they can recreate something they easily have thousands of examples of in their training data. There's probably a better base RTS on github somewhere for free.
Well, it is a silly test, not a scientific benchmark.
However, I would say it is a measure (not the measure). If you look at the entries, there's a lot of variation - definitely not something they memorized outright.
And the test itself is deceptively simple. You need to do canvas rendering, there's pathfinding, command queueing, terrain generation, etc. There are some subtle click handler bugs (various LLMs often stumble on those). And I ask the model to do it all in one file, further increasing the complexity of the task.
And the result is something that you can instantly evaluate. And if the result is any good, even play! So yeah, I think it's a fair test.
I'm sure it'll get saturated at some point. Actually I started with Minesweeper and switched to RTS last December, because Minesweeper was being saturated. I'm expecting (hoping?) the RTS test will last until the end of this year...
Ask it to generate a completely bananas game idea and it will easily do it. Prototyping any type of game with simple graphics has been solved since many generations of models back. Claiming it's because it has that game in its training data is nonsense.
I really wish there was just an easy guide on when to use Sol vs Terra vs Luna, and it just moves further into confusing territory when it comes to naming.
The naming convention is especially difficult to decipher depending on what your native language is. Of course a latin language speaker might be able to easily determine oh yeah each one is slightly bigger than the other but I still think it borderlines too confusing.
That aside all the numbers look amazing, and I'll be happy to probably main this alongside grok-4.5 for a while comparing the two on price and efficiency.
I vastly prefer the direction that OpenAI seems to be going with token efficiency and performance compared to Anthropic who seems to be moving towards a world where you just token-max as much as possible ignoring any and all costs.
Use Luna. It's more performant than 5.5 and it's cheap. Hopefully it's cheap because it's more environmentally friendly than the bigger models. So you're doing a good thing. If it's a smaller model it may even be faster, but I haven't looked into it yet.
My guess is that it's the same for Haiku/Sonnet/Opus: Biggest model for architecture and high level planning and technically challenging problems, medium model for simple implementation tasks, small model is for nothing
Previously it was much more obvious which model to reach for depending on your use case because they had the mini and nano naming conventions.
Getting rid of that seems like a step back. Just a personal nit though.
I've seen buzz about this elsewhere as well but to me effort levels seem more like spend limits disguised with another word. I don't think they should even exist.
The naming convention is bizarre and doesn't really mean anything to normies. Trying to pick between "Sol" and "Terra" is like asking the average person if they want the Max or the Ultra chip.
That isn't what "genuinely asking" looks like, you're criticizing using "questions" as cover. It isn't subtle, nor is it constructive.
I agree with them, Sol, Terra, and Luna are confusing names. They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
It feels like it was adding by the marketing department.
I'd agree it is similar to Anthropic's naming scheme, which I'd argue shares the same problems as this. It improves marketability/googlability, but decreases actual comprehension.
You don't actually explain why or how these names are "easy to understand" just state that they simply are. That's great; to me, they aren't obvious or intuitive at all. May have well just start randomly pointing at dictionary words.
I collect roman coins, with latin legends, so the sun/earth/moon references jumped out at me, and partly based on the opus/sonnet/haiku precedent I assumed that these names were referring to different model sizes/prices in a way that mapped to the names (Sun > Earth > Moon).
I'll admit though that until recently I never really thought about Anthropic's naming scheme as having meaning (an Opus being longer than a Sonnet, being longer than a Haiku).
>They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
But do they though? When do you use GPT-5.6-Max-Low vs. GPT-5.6-Plus High? Or GPT-5.6-Fast-Xhigh? What's the Pareto optimal choice (outcome and price)? According to the benches it seems to bop around and the even if the benches are accurate the best choice isn't always consistent.
> When do you use GPT-5.6-Max-Low vs. GPT-5.6-Plus High?
You don't, because that isn't something I proposed using for model naming.
I called them GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast. Reasoning levels are distinct from the model design itself, and the UI makes that clear.
Plus, using that same flawed argument this would be called GPT-5.6-Sol-Low or GPT-5.6-Luna-High which also makes no sense/is confusing. So that argument applies (or more accurately doesn't), no matter the model names.
I do know what Sol/Terra/Luna mean, but was also confused for a second on the hierarchy. After doing a bit or research it dawned on me that they are arranged in the order of the sizes of the celestial objects but it somehow wasn't immediately obvious to me from the start.
Anthropic ships models with a helpful one-liner tag that makes the model hierarchy obvious. I think it wouldn't hurt if OpenAI did the same.
Did you not read the second sentence? Obviously I know what sol is given my first language being Spanish. I'm just speaking in a general sense that it can be confusing for others.
I already know plenty who had no clue what the difference between Terra and Luna would be.
My first instinct was Sol > Luna > Terra, since Sol is the farthest away, then Luna, and Terra is the closest. Size was not my first instinct. Or should Terra be the best model because its closest to people, then Luna because there have been people on it, then Sol be the worst because no human has been there?
The frontier graph on all these benchmark are extremely in favor of 5.6 Sol over Fable, more than the best model comparisons in previous iterations.
I'd like to know how cherry-picked this is, and what tests it performed less overwhelmingly in, but I suppose that info is not going to be on this post.
If it pans out to be as good as it says, that's great. On the other hand, if this model is not overwhelmingly impressive over Fable, I will lose what remaining trust I had in these announcements.
They do disclose that they scored much lower than Fable on SWEBench Pro, which is a pretty high-quality benchmark. I think it's partially just about what they choose to emphasize...
People will downvote you because this comment is "not appropriate" for HN, but there were countless conversations on HN about how important these benchmarks are.
There has been a lot of chatter ever since the Mythos scores had been release that SWEbench pro had major contamination and that Mythos had memorized many questions that lacked the context to be solvable on their own. And now with OpenAI saying a large number of the questions are broken, I think it's worth taking that single outlier benchmark with some salt when the overall trend is that 5.6 is very competitive with Mythos at about half the price.
I totally missed that, because in the charts they showcase for coding, the SWEBench score is not present, they only include it at the end of the post in tables. Hmm.
The SWEBench benchmarks are really gamed at this point and should not be trusted period. The solutions are effectively in the training sets and have been for a while.
Then we are left with what? FrontierCode maybe? IIRC that one evaluates not only if tests pass but also code quality - e.g. whether the maintainer would accept the pull request as is.
The charts are also extremely difficult to parse. They seem auto-generated. Dataset coloring is atrocious.
Regarding your main point, yes, I agree. My impression (as someone who uses both Codex and Claude Code daily) is that OpenAI does a fair amount of benchmaxxing.
Anyone else noticed the "Extended: Fable 5 is included in your weekly limit
through July 12 blablabla" disappeared from claude code? Did they panic-delete the july 12th deadline ?
100. I'm so tired of being treated like a drug addict by them. I'm currently sitting on _four_ "reset vouchers" from OpenAI so I get basically all week to play with 5.6 Sol to my hearts content. The amount of positive sentiment that brings to me should really be a concern for Anthropic who is increasingly alienating me away with their shit strategies.
Or if you want to see some in 3D, OpenAI featured a pelican riding a tricycle, bicycle, pony and another pelican in their livestream this morning: https://www.youtube.com/live/Wq45rvPGNHs?t=1070s
On the one hand: yes, pelicans on bikes are definitely in the training set at this point.
On the other hand: the test is clearly not saturated, given that you can see a clear difference in output at the various reasoning levels / model versions.
I sort of agree, but within the same model I expect the reasoning effort to be reflected in the quality of output and that's basically how it played out. When you're comparing different models, then it's just who benchmaxxed the best and there's not a lot of value there.
I don't know. If they were training on this, I feel like they would be able to get the shape of a bike frame right; it's a pretty simple polygon, and a lot of the bike frames that are getting generated would be impossible to steer.
I think the 'pelican test' is becoming useless. It's been around long enough that now I'm sure good examples are in the training data, and hell they might even do some hand tuning to make it do a decent job since they know people will ask about it.
But either way, with no real way to visualize the result of the text it starts with - it will always be stabbing in the dark. It can't understand conceptually what any of it should look like and then refine the SVG to improve it gradually. It just throws darts at a wall and hopes it comes out alright.
Honestly that's the main value I get from it myself - making a pelican means I have to figure out API keys and how to talk to the provider, or how to run it locally for the local models.
Pelicans, maybe, but the point is to measure how good the "internal visualization" abilities are. Throw curveballs, like a unicorn with a duck bill serving coffee at a basketball court. An elephant playing a piano while its trunk swings a baseball bat at a tiny alien spaceship buzzing its head.
Have them use tikz instead of svg, or have it write code that moves the cursor and draws the thing in paint.
Compositionality and visualization are generally much, much better at each new generation / release cycle.
It's fascinating how well models have internalized visualizing things without actually having joint embeddings / broad multimodality.
Cool. I still find these a useful visualization of some the qualities of llms. Even if they did train for [animal] on [vehicle] svg, it's still nice to see at a glance how the different models and reasoning levels perform. Lunar misses part of the frame, except on max reasoning. While most of the others have a mostly correct bike at all reasoning levels.
I once used something like karpathy's auto-scientist to mutate the prompts and rank them with a vison model. Some of the winners where pretty neat. I think they have a lot more style than the gpt-5.6 ones. https://xcancel.com/xundecidability/status/20449185674144196...
What's strange with this is the prompt "Photorealistic photograph of a pelican riding a bicycle down a coastal boardwalk, wings gripping the handlebars, webbed feet on the pedals, large orange bill, detailed feather texture, golden hour lighting, shallow depth of field, shot on a DSLR with 85mm lens, natural motion blur on the wheels" produced, well, exactly what I asked it for. I wonder if I tell it then to make it SVG ...
I'm waiting for the day that the "generate a Pelican" test comes back with a SVG-art like illustration of a Pelican equipment case, like a model 1620 or similar.
I haven't tried this in a few months, but last time I tried a loop that rendered the pelican and asked for improvements the results were actually quite disappointing. Be interesting to try that again against GPT-5.6 at Claude Fable 5 though.
They said in the AI community, a pelican riding a bicycle is a good test to measure effectiveness of the model, wondering if they were referring to you, or is it really a standard in the AI community ?
Also would be good to have a tool where users can select models and instantly see each model's generated pelicans. That will make it easy to compare the output of different models.
Simon did start the pelicans on bicycles as an SVG, but I think it's more of a fun goofy thing to see how the model performs at. I don't think it has a direct correlation to a model's performance though.
I eagerly await the models replying with that: "I'd be happy to create a pelican riding a bicycle, but just a note that this might already be in my training data. Simon."
I think all of Luna's are bad. The only decent one is sol @ xhigh. Even sol @ max is weird. Sol @ high and @ medium are ok, and every other single one across every model is bad.
Strong disagree, but to each their own. For sol I really like how only medium uses the wings on the handlebars to ride the bike. For all the other sols the pelican evolved a new set of arms separate from the wings.
Things I have been struggling with Fable over and GPT 5.5, were just solved handily by SOL in a real "thank you, next problem" kind of way. Overall, something that just works is way less wasteful for your usage than struggling back and forth for hours.
Unfortunately, I'm finding that in long-form agentic use, when I'm trying to use Sol, I keep tripping guardrails – moreso than even Fable, somehow.
I don't know exactly what part of my codebase is triggering it, so I'm going to have to keep poking, but apparently the guardrails are not that gentle despite the phrasing. :(
>> approximately 700,000 A100e GPU hours of black-box automated red teaming
Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.
Not nothing either, but far less astounding sounding than 700k hrs.
Wait, what do you mean? 700k A100e hours are equal to 200 hours of a GB300 NVL72 rack? One GB300 NVL72, 72-GPU rack has equal processing power to 3500 A100e GPUs?
"GPT‑5.6 delivers a step change in design judgment. With only high-level direction, GPT‑5.6 creates tasteful, ergonomic, and functional interfaces. Its stronger computer-use capabilities let it inspect and refine the rendered result—not just generate the underlying code or content—so it can catch visual and functional issues and apply finishing touches before handing the work back."
This one is really promising, as it may allow to close major gap with Claude in design/UI skills
Computer-use is a big limitation that my 2015 Macbook Pro cannot handle. I find the Codex cli says it looks at the end output artifact but so often it fails to refine it into acceptable form. If it could use my computer screen and visual inputs for review, it might be able to actually design documents/powerpoints/etc. I'm juicing everything I can out of the 11 year old laptop and I'm honestly impressed at what it can still do.
+1. I've been only using Sonnet/Opus these days for UI work because GPT 5.5 just can't do any of that. Its just really terrible. Eager to give this one a try.
Agreed, I’m looking forward to trying it out. I think that the rise of visual design skills that are pretty clearly targeted towards Codex users has lit a bit of a fire under their butts.
GPT usually performs better on DeepSWE while Claude does better on FrontierCode. These two coding benchmarks are pretty much the only ones right now that's still worth taking a look at imo.
DeepSWE seems to strongly, strongly prefer ChatGPT models. There were also major flaws in its methodology pointed out recently, that overlap strongly with the flaws OpenAI pointed out in its SWE Verified report.
I use both ChatGPT and Claude for engineering work on a daily basis, touching performance critical code to application backends to frontend work, and I've found that DeepSWE scores don't reflect my reality when I assess high quality output from the models/harnesses.
Not that Opus always beats GPT 5.5., but that 5.5 is ahead of Opus on a general benchmark smells off to me.
I flip back and forth between whoever currently has the more powerful frontier model that isn't cost prohibitive - subscriptions only, API pricing a non-starter. Today that's Fable 5 which has been excellent, as soon as it's Sol I'll switch to that. The OAI/Anthropic harness behavior has mostly stabilized for me with consistent AGENTS.md that I sync with CLAUDE.md - I like pi (pi.dev) and have tried to build it up to get performance comparable to the two "first-party" harnesses, I'm just not there yet.
One major sticking criteria for not going with OpenCode / pi for all of my coding is I want access to the tier-1 frontier model of the day without API pricing - e.g. afaik I can't use Fable 5 via pi harness even though I have a subscription, so for this week I'm on Claude Code. It's not the need to Fable 5 for everything, but even if I just want the marginal intelligence benefit to stress test an architecture decision, it's a safety blanket to know there isn't a ~smarter~ model I could have used. And for my use cases, the doggedness and capability of these frontier models has been insanely effective.
My feeling is we're still in the Uber era subsidy period - the moment the subscriptions either try to lock me in longer than a month or stop OAI/Anthropic stop delivering frontier models in the subscriptions, I'm out - switching fully over to pi.dev or another OS harness and routing my token spend via OpenRouter or offloading to Qwen locally. Then I'll have to put an accurate dollar amount on frontier intelligence.
> My feeling is we're still in the Uber era subsidy period
I often wonder whether this doesn't continue indefinitely.
Uber was able to do this because it was just them and Lyft playing second fiddle, with a huge barrier to entry once the network effects had kicked in.
It just seems like the model space has way too many competitors, + OSS/Local options for them to ever be able to jack up their prices. At least once the datacenter bottleneck has been cleared.
I'm working on a multi-harness IDE that supports custom agent workflows and skills that are shared between any harnesses it wraps over. I think it might prove handy for a workflow like yours.
Interesting that Sol (low) did better than Sol (medium) in your benchmark (and is barely more expensive than Terra). I too have been using 5.3 codex as a cheap-but-good model and are switching to Terra (xhigh).
Yeah, for some reason the (low) versions do really well, like they think directly of the solution instead of going around all the edge-cases and getting lost in one of them.
Benchmarks look really promising. Suspiciously good, even. I guess we’ll see soon enough.
My question to previewers: how are the guardrails for random joe that wasn’t personally blessed by the ai pope to access the non-nerfed model? Fable is a nightmare in this regard, but I’m not sure whether 5.6 also gets a critical side-eye from the gubmint when you ask it to fix bugs in your code (you filthy hacker, you).
I almost immediately ran into "This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a
quicker response, though it may be less capable of handling complex requests."
Which is something I've never seen with codex before, and I wasn't doing anything funky. Just writing CUDA kernels and benchmarks for them.
Is it actually usable though? Because the Fable situation is just obnoxious. If OpenAI's Fable equivalent is actually usable, I'll cancel my Anthropic subscription on the spot.
Yes it's usable, however it does make it difficult to just do /goal and walk away if this is happening periodically. I don't do that often, but for well defined things I sometimes do.
Seeing how Anthropomorphic just reset usage quotas back to 0 and the other day extended Fable sub inclusion by a few days, I have a feeling they might not drop Fable out of sub after all, because like you I would most definitely take a long good look at codex at that point.
It's not just the API pricing either, there's also the constant uncertainty. They pull the model then put it back up, they say the model is going away then suddenly it's not. And then there's the fact Fable is barely usable because it randomly downgrades to Opus out of nowhere whenever it thinks about exploits.
It's definitely good that Anthropic's feeling the pressure. Anthropic has worn out their welcome with this "safety" nonsense. If OpenAI actually lets me use the LLMs on a subscription without any of this bullshit, I'll definitely switch.
Based on the Intelligence vs. Cost graph, not clear to me why anyone would use Terra? Luna looks quite interesting though, happy to see OpenAI still serving the more budget-oriented side of the market (seems like Anthropic and Google have lost interest there).
Very interesting: I wonder if the RL approach is diverging between Anthropic and OAI?
I noticed that Fable uses shell tools almost exclusively (even to search and edit files), compared to previous Anthropic models.
Having run some experiments with 5.6, I notice that it uses built-in file systems and provider native tools much more (not shell tools), compared to previous OAI models.
I can't try it since it hasn't appeared in my Codex yet, but this is is necessary from OpenAI in my opinion. Fable is just so much better at understanding broad context. I only use GPT 5.5 for straight forward easy to describe tasks, and it does crush those. But I spend a lot more time steering Codex towards good design on broad concept type tasks, ones that Fable shows sometimes surprising clarity.
I look forward to seeing how it compares once I have access. Not getting tripped by spurious safe guard flags could be an advantage.
I’m interested in knowing how each of GPT 5.6’s variants fare in non-English writing/translation tasks.
GPT 5.5 has a tendency to write English calques and non-idiomatic prose in other languages. Although that can be somewhat tamed with detailed instructions and a corpus of confusing terms, the model’s output often reads like a literal translation rather than native prose. Since I notice these issues most clearly in languages I know well, it makes me reluctant to trust the model’s output in languages in which I’m less proficient.
Ironically, ChatGPT began as a simple text-generation tool, but much of its offerings and benchmarks now focus on coding and agentic workflows, while leaving behind what made it notable in the first place.
Just used terra ultra for exactly one prompt in codex and it ate through my full 5h window in about 10mns (20$ plan). The results look pretty good though. Luckily I have had my chatGPT subscription for a while and have a bunch of resets available (nice compared to anthropic).
Assuming I take the 5x plan it would give me about an hour of active sessions with terra ultra (maybe ultra is not good value regarding tokens?), not even using Sol yet. Does everyone using codex use the 200$ plan?
I normally use the 100$ anthropic plan and barely ever reach the usage limit.
Well, yes, as explicitly stated on https://openai.com/index/gpt-5-6/: "ultra goes further by coordinating four agents in parallel by default, trading higher token use for stronger results and faster time-to-result on demanding tasks."
I use the $20 plan, but I don't code all day every day.
With Codex, it is my experience that I can churn through a 5h window in no time with newer models -- especially when they're new. So I tend to use fancier models for planning, and the less-fancy models for writing code based on that plan. I switch to the fanciest model if any part of this gets stuck.
If I've got a something big-ish to work on, I pay attention to the reset timers so I can get more of it done in one chunk.
Models seem to slowly get better/relatively less-expensive as they age. (It isn't clear to me if that's because the cost actually goes down, or if the allotment goes up, or if things get more efficient in unseen ways, or what. OpenAI is vague AF about what we get for the $20 that we pay.)
Is any of those comparisons about Pro vs non-Pro (Pro is only available in $100+ plans)? I am curious about that but I think Sol, Terra, Luna are different sizes of it without the Pro part, and I want to know how much worse do I have it on the $20 plan compared to if I upgrade.
Huh, a good alternative just as anthropic's 50% weekly subscription subsidy is ending this weekend. Time to see if it's benchmaxxed or actually a strong leap over GPT5.5.
They also seem to really not care about alignment, or care about it in the wrong way. It's entirely missing in the blogpost and there are some concerning bits in the model card, seemingly treating CoT controllability as something to be "investigated" rather than the warning sign it's supposed to be.
it seems terra is pretty much useless, you either want luna max for everyday coding (cheaper and same perf as 5.5 high), or sol xhigh/max for demanding tasks
Parameter: reasoning_effort
Function tools with reasoning_effort are not supported for gpt-5.6-sol in /v1/chat/completions.
To use function tools, use /v1/responses or set reasoning_effort to 'none'.'
Official OAI .NET library. Even when I override the currently experimental [?] flag to 'none', it will still occasionally throw this error (about 5% of the time).
I hope we aren't trying to push customers off the chat completion endpoint... Responses endpoint looks great on paper, but the business wants more visibility and control over the reasoning process than this product currently offers.
For context, I have access to MS Copilot through my workplace. To see what it looks like, I have tried to login through https://copilot.microsoft.com/ , where I was informed that my account, although recognised, is not yet supported. However, I can get more or less the same chat window, with access to all the data, through https://m365.cloud.microsoft/ A redirect could have been useful.
I used to pride myself on not being the "fonts too pointy, scroll too buttery" crowd! But AI has brought me full circle and now nothing removes my interest in reading even a single word on a page faster than purple gradient greeble-afflicted tailwind-slop models put out without stronger prompting/references
One of my best use cases for the short duration I have fable is to use it to create the plan and acceptance test files then use GPT 5.5 Pro to do an adversarial review on the plan then feed that feedback into fable to fix the plan.
Zero information on the knowledge cutoff. The model itself responds it's June 2024 which is weird given that GPT-5.5 has knowledge cutoff at August 2025.
> Instead of requiring developers to script every step or passing every tool response back through the model, Programmatic Tool Calling in the Responses API can filter large amounts of intermediate data, retain only what matters, and adapt its workflow along the way.
There is an issue on the page that causes the benchmark tables to get cut off. If you highlight and drag right you can see a few more models like Gemini and Claude Opus. It's also interesting that they introduced explicit caching, which is something that only Anthropic had for a long time.
Looks like I have access to gpt-5.6-terra and luna. How does one decide between gpt-5.5 and gpt-5.6-terra? Pricing is similar, but it's hard to tell if it's better..
this is exactly my question. I would expect that luna is analogous to mini before, but is terra equivalent/better than 5.5 and Sol is a step above? or is terra nerfed and 5.5 is analogous to sol?
I wish model launches were like proper product releases
it's impossible to _try_ it out on release!
it's not on their codex subscription, or the web/mobile chatgpt interfaces, or aws bedrock, etc. I just cant find a working endpoint with the latest model after they announce
The announcement says they're rolling it out over the next 24 hours or so. I think it's reasonable to do a slow-roll-out release for one of the most used products on the internet.
Looks like a great set of models, but there are about 20 different thinking/model levels here in this family and they are very complex to pick the right one for the task
E.g. for GeneBench Pro, it looks like you would always use GPT-5.6 Sol over Terra/Luna, its pareto optimal.
For Agents Last Exam, you would maybe want Luna, then Terra, then Luna, then Sol as you increasingly budget for tasks.
I feel that there may need to be a new auto mode in many of these cases. It selects the best model and thinking given a particular problem.
Feels like it's going to have to go that way eventually, because here we have about 20 different model and thinking levels you could use, and they're not obvious which ones are right for the given use case.
I think 5.6 Sol is only as good as 5.5 or Opus 4.8 in terms of getting its given work done. It just has an uncanny ability to pickup more work that it can tackle next that the older models lack, or have not been trained to do before.
Where folks are seeing a difference between working with Fable or 5.6 I think also boils down to this phase shift.
Not specific to OpenAI / Codex, but I'm curious what people are doing to protect themselves from any destructive actions by their coding agents? Just install and pray? Explicity approve all actions? Reconfigure for safety? Run in a sandbox (Docker) ?
Typically I just want to isolate the agent disallowing it from accessing other parts of the filesystem. Using a different user might be enough, but I typically use [bubblewrap](https://github.com/containers/bubblewrap).
More seriously, I was blindly trusting the auto-classifier from claude code (same as the middle option when you do `/permissions` in codex), and it actually allowed the agent to do pretty hardcore `rm` and `git push --force-with-lease` commands, which I would have expected to have to approve manually. Luckily no major issue from those yet.
The best option imo is the integrated cloud environments from claude code (not sure yet if there's a codex equivalent). It spawns a VM in the cloud where the agent runs, and you can open a PR from the app when it's done. Very smooth experience
Interesting - I'd never heard of this Claude Code VM option.
Does it auto install all the dev/test tools it needs, maybe including things like web server & browser? Does your code live in the VM, or in some external repository? Is the lifetime of the VM the same as the agent, or does it persist until you remove it?
I use the auto-reviewer for actions outside the builtin sandbox.
So far this has been rock solid, and tens of millions of developers use this setup without issue.
It is not going to wipe our hard disks. At least I hope so. Fable and GPT 5.6 have been ever more proactive, and GPT 5.6 is automating the AppStore on my machine to download an Xcode update while I am typing this.
Yes. In Codex it is called 'Approve for me', in Claude it is 'Auto mode'.
I believe in both cases it is prompting a model with a fresh context that is tasked with reviewing the reason for the action.
With Claude, I have seen that if the reviewer does reject the proposed action, it responds with a long text about how the Agent should not try to work around this rejection, and instead prompt the user for an explicit approval of the proposed action.
I still just explicitly approve all actions and review all code (unless it's a personal/throwaway project no one else will ever touch/use/see). I know a lot of people that run in a sandbox though. That said, I'm sure there are lots of people that just yolo it and hope for the best.
What destructive actions are you afraid of in particular? Honestly the models are pretty smart, I let the agents go --yolo and nothing bad has ever happened (yet) that couldn't be solved with git.
I'm not concerned about the code it's working on, but rather anything else - modifying files outside of the project dir (e.g. incorrect tool call), modifying system configuration, doing something bad on the internet, etc.
It is a valid concern but I've been running yolo mode since the inception, using Claude and now Codex. I'm not bragging or anything, since I'm feeding my own curiosity too, trying to answer what's the worst that could've happened? So far, nothing catastrophic that I recall.
Oh man, I love capitalism spoiling us here. I was just enjoying my extra Fable credits, now I'll switch to using 5.6 this weekend. I was planning to ration my Anthropic credits, I guess now I do not have to. And I was half wondering if exactly this would happen: right when Fable usage credits were starting to kick in for people, OAI swoops in and takes the puck. As much the AI craze is crazy, this play by play part is pretty fun.
I never have have the issues most people talk about ... I feel like most were never Devs before ai and don't know what they actually need done when prompting. that on top of not utilizing good tools such as a codebase indexer, lsp and a project scaffold.
Yeah, I pretty much had to switch to using GPT rather than Opus completely for all my security benchmarking and harness development. I was annoyed enough to blog about it: https://swelljoe.com/post/why-i-had-to-switch-to-gpt/
From what I see on ArtificialAnalysis.ai, OpenAI cranked up the reasoning effort to surpass Fable, and now the cost/task for 5.6 max is closer to the Claude Opus 4.7 and 4.8 cost/task.
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
My Codex app got upgraded to the new unified ChatGPT app. I don't see Sol available though. Only Terra and Luna. I'm on the Pro plan. Anyone else see it?
Its an extremely capable model. I think the way we need to approach works shifts again. We need to get our harnesses/workflows to let it gather some momentum on the first couple rounds but then we also need to structure it so that it can slingshot and accomplish the long range goal.
I assume they're jealous of the Fable/Mythos hype. People talk about Fable like it's a whole new thing, rather than another incremental improvement over the existing best models (which has happened several times and continues to happen).
Overloaded in Codex, no indication if it is already in ChatGPT and I can't use it in the API even though it says it should be available. Typical horrible OpenAI launch. Glad that Anthropic just reset the rate limits so I will go back to Fable again.
GPT 5.6 Sol is a token hog. After implementing the task, it started some "reviews" I didn't ask for - they consumed 19.5M and 11.9M tokens, while the task itself was below 5M tokens.
> GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints (opens in a new window) and a 30-minute minimum cache life.
Great to read they are moving away from the 5 minute cache defaults. Hopefully other providers follow soon!
Almost immediately ran into some the kind of gatekeeping I've heard Claude Code users complaining about with Fable. Not sure why, I just had it working on writing benchmarks for some CUDA kernels. Nothing security related:
"This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests."
At least it gave me the option of waiting instead of just unceremoniously downgrading me. Appears to be making progress but... weird?
I think the most interesting part of this is that OpenAI is going way easier on the classifiers than Anthropic. They explicitly state that many defensive cybersecurity uses are supported and implicitly criticize Anthropic's stance on Fable's uses by saying that overblocking cyber requests is itself a major security risk as more AI models continue to advance in intelligence. I have so many questions as to what is going on on a game theoretic level in the AI space in the past two months, it seems like multiple actors have realized their incentives are really quite different than they originally thought.
It's good to see labs taking into account the cost/task.
Grok 4.5 is interesting because it's smart enough at great price. It seems gpt 5.6 is right there with great efficiency and great pricing.
Working with Fable has been a great experience, but at the end of the day, if you can get only 10% of your work done because it just burns through tokens, that's not that interesting.
I've been mostly using Opus and Fable high for planning and codex 5.5 medium for implementations. Claude is also the only model i can use for design tasks. If gpt 5.6 can finally deliver on the design side, it might be time to ditch the Claude sub and go full Gpt.
For writing GPT which i was subscribed to Fall 2024 to March 2026 (laid off) is superior to Gemini. Been using Gemini since March mostly and they offered a $10 a month plan so i took it. Though today realizing GPT is superior to help me write I am back to being a paying customer. Im in full swing mode to get back into the job market (get the heck away from UI/UX which is now a stupid career in terms of number of jobs out there and in the future there will continue to be less) pivoting into product management (can vibe code anything now) and or customer relations. Hopefully GPT helps me with this pivot and Im again gainfully employed!
GPT Terra is 50% cheaper than 5.5 while being more performant. So it’s like a straight up 50% reduction in cost!
That leads me to a question. Why wouldn’t they just default to terra in ChatGPT in the last few months? If they didn’t then they burnt money for no reason by giving a shittier model at a higher price
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
I'd say answer , the opus is no longer undisputed. grok + gpt models are very competitive + glm if you are ok to wait 3-4 times longer, unless you have some unique access to GPU
"On Agents’ Last Exam (opens in a new window), an evaluation of long-running professional workflows across 55 fields, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT‑5.6 Terra and GPT‑5.6 Luna outperform Fable 5 at around one-sixteenth the cost. "
Some pretty big claims and results! Excited to see how it feels during usage.
I use Fable and 5.5 extensively and I still find both have a place in my toolkit, i.e. Fable IS good but it isn't perfect, and it's still better to play them off against each other. I have Fable and 5.5 write plans and have them adversarially review each other's plans.
Having this amount of competition in the coding model space is good for all of us.
I think this is the phase shift 5.6 (Sol set to Ultra) is bringing to the table. Until now we have become accustomed to asking models to continue and their natural inclination is always to stop.
Now OpenAI have flipped it around and for the first time are asking us to steer or stop the model instead, and its own inclination is to keep going. We now have to decide when we need to steer or want to catch up on our understanding of the work done but it will keep going.
Here's me using a Gemini chat log scraper (from Gdrive) then dumping my prompt+Gemini response into local AI
Never go over the free limits in Gemini Pro.
Gemini is great at research and architecture, and my 30 years experience in programming everything; for fun or work; means together there is little to no code slop.
Add to project repo some git submodules of reference source code; boom, bobs your uncle
Zero reason to sign up for OAI or Claude. With employers realizing the costs are more than employees, local models getting more powerful, and models in chips just a few years out, neither of the one note LLM companies without diversified services and R&D portfolios gonna last
i'm not happy with how openai is trying to pit 5.6 sol as a cheaper equivalent to fable here
for one thing, they said that on AA, sol is "within one point of fable" at 58.9 vs 59.9 but don't clarify that the latter is with safeguards where ~8% of the tasks got routed to opus
i'm not rooting for either and genuinely think that the token efficiency and cheaper price are important but this sort of thing just feels disingenuous :-/
As usual, even though GPT-5.6 is releasing today, the rollout in ChatGPT and Codex will be gradual over many hours so that we can make sure service remains stable for everyone (same as our previous launches). We usually start with Pro/Enterprise accounts and then work our way down to Plus. We know it's slightly annoying to have to wait a random amount of time, but we do it this way to keep service maximally stable.
The timescale is typically hours not minutes, so if you don't see it now, I'd try again later today.
We mention it will be a gradual rollout over the next 24 hours in the Availability section at the bottom of the blog but I admit it's pretty buried.
Is this bug fixed with 5.6? If not, it probably doesn’t matter which version Codex users are getting because the overall result is dramatically worse than stated by Open AI advertising: https://github.com/openai/codex/issues/30364
because they're stealing from the frontier models. they're gaming the benchmarks. look how bad glm 5.2 is on cursors evals. gmhit garbage , but it gets glazed as God tier.
SWE-Bench pro is pretty much useless now even though many ppl still look at it. OpenAI published a report yesterday saying so as well. Only look at DeepSWE and FrontierCode right now for coding imo.
SWE-bench series just aren't that great by today's standard, even Anthropic previously stated Claude had memorized solutions for the non Pro version of the benchmark, I suspect the recent increase in the score for the Pro version probably also had similar behaviors.
But anyway, I think it's pretty useless to look at SWE Bench's now when other way better benchmarks exist.
And they'd be right, it's an almost saturated benchmark where even some subpar open source models score very well on. And most models are clustered within a small range so it really doesn't tell you much.
> On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol with max reasoning sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less.
> That advantage extends across the family: Terra performs just above Fable 5, while Luna outperforms Opus 4.8; each does so in roughly one-third of the time, with about half as many output tokens, and at approximately one-quarter the estimated cost.
Wow. I don't believe it. Every indication and twitter post told me that Fable is much more intelligent than Sol and here we are told that even Terra outperforms Fable?
Not only that, Sol doesn't even come with run time classifiers. So it is even more suspicious.
What's even stranger is that OpenAI is directly referencing a competitor in this direct way.
Holy shit. They must be feeling very threatened by Fable if they're spending this much energy talking about it in the release notes for their own model.
> GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output.
Just as expensive as Fable 5. But of course, another slot machine upgrade but the costs will keep going up and the open weight models from china will continue to race everyone else to $0.
Looking forward to the next version of GLM, Qwen, Deepseek and Minimax.
Also watching deepseek closely. Seems like US frontier labs only know how to throw money at things as opposed to actually make smart improvements to the algorithms.
The marketing team must've done research that said "people are starting to think that you guys are evil-water-stealing-lay-off-loving-bubble-bursting scumbags" and decided to really lean into the small family business and happy font vibes!
All of them closely collaborate with the government. LLMs are a national security priority and are vetted. Claude AI was used by Palantir's Maven to target the Minab school that led to a triple tap strike killing over 150 schoolchildren.
Claude makes suggestions for targets and humans review and approve them. We always knew there was a human in the process but if there's one massive takeaway from years of AI ethics research, it's that there is a very clear and well-documented human bias towards automated answers when there's any ambiguity.
Including a human in the loop does not excuse the fact that AI was trusted in a process that decides who lives and dies.
Weirdly, normally new ChatGPT releases are head and shoulders above anything else, but according to OpenAI's own evaluation, Anthropic's Mythos outperforms ChatGPT in quite a few benchmarks: https://openai.com/index/gpt-5-6/.
ChatGPT 6 must be deep in the pipeline and will be released within the next few months. Maybe that's why this release is versioned 5.6, not 6.0.
The developer's guide (https://developers.openai.com/api/docs/guides/latest-model) has some interesting semantic tips for using the model:
> Intent understanding: GPT-5.6 can better infer the user’s underlying goal and intended level of work without you specifying every step. Continue to state important constraints, approval boundaries, and success criteria explicitly.
> Original image detail: GPT-5.6 preserves the original dimensions of images sent with original or auto detail instead of resizing them to a patch budget or pixel-dimension limit.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Control warmth: GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic.
> ...tips for using the model:
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
I don't follow. Isn't "the model actually cares and will do what you say" a reason to use those kinds of instructions more liberally?
I think they’re saying it’s irrelevant now, possibly because it’s less likely to trail off on meandering thought bubbles.
> Avoid generic brevity instructions
That part is confusing because it's not like they provide an example of how default GPT-5.6 output compares with GPT-5.5 both with default output and prompted for brevity. Whenever I use such prompts, it's usually because I want the model to give me the gist in a few sentences. I'd be stunned if GPT-5.6 was that concise by default. I would think that could "break" a lot of things for developers who didn't know to make prompt changes after upgrading to 5.6. What if you were expecting GPT to be as wordy as it usually is? Then suddenly your output is not wordy enough?
Smells like OpenAI trying its best to stave off financial armageddon for another few months. Then again, I'm not sure why they chose to waste so much output computation on verbal diarrhea all this time up to now.
If you conceptualize this as “there is an appropriate amount of brevity for each situation” then it would be expected for a better model to use different amounts of brevity if it gets better at determining the appropriate amount.
My view is that popular models by default output wildly excessive amounts of prose for nearly every use case, so if this changes in a new model that’s a pure win.
The models don't get better, except when a new one is released. Their performance depends solely on the model training before release and how well you curate the context you feed it. That's it. Contrary to popular belief these things are not intelligent.
>The models don't get better, except when a new one is released.
My brother in Christ this entire thread is talking about the new model that was released
This has absolutely nothing to do with the comment you replied to.
> could "break" a lot of things for developers who didn't know to make prompt changes after upgrading to 5.6.
How does this differ from the other changes in behavior in 5.6 that will also break things? New models always break things.
It seems like the way brevity instructions have changed is mis-aligned with how most people would expect to use them or are currently using them.
Here's the example they give:
> Instead of asking for the shortest possible answer, replace brevity instructions with prioritization:
> Lead with the conclusion. Include the evidence needed to support it, any material caveat, and the next action. Omit secondary detail and repetition.
> Keep all required facts, decisions, caveats, and next steps. Trim introductions, repetition, generic reassurance, and optional background first.
Generally speaking, when I ask for a short answer, I want a short answer because I'm not really willing to read through a bunch of bullshit to get to a summary. Putting the onus back on me to assume what the model will return and write a longer prompt detailing exactly what information I want completely misses the point of why I'm asking for a short answer in the first place.
> Lead with conclusion.
I would presume (perhaps falsely?) that an instruction like this would lead to the model presenting a conclusion not supported by the evidence, and potentially backtracking as it then tries to justify said conclusion.
Yes, if deliberation happens, the model should figure out what it wants to say during that phase; but if you're using auto mode, the model is not going to be doing any deliberating half the time. In those cases, the output blathering is the model's only chance for deliberation. It "thinks as it talks", per se.
Given that, I would advise a different approach: let it blather, but then get it to write you a conclusion at the end that the model can guarantee will obviate the need to read any of the blathering.
I.e. advise the model to add an "executive summary" to the end of any non-trivial-in-length response. With some wording to carefully navigate the model between "the summary is itself too long" vs "the summary acts more like clickbait, leaving out necessary detail such that it requires actually reading the blather."
Not sure exactly what that wording would look like. I imagine something like "write your postscript executive summary as if you were a senior CIA intelligence analyst summarizing ground-level reports into a daily digest for the Joint Chiefs of Staff. Take up as little of their time as possible, but ensure that any detail critical to decision-making is retained." (But that phrasing might only be useful if the model is delivering a certain type of response, and actively counter-productive otherwise. This kind of thing is delicate.)
You are absolutely correct. The second suboptimal part of the prompt is this:
> Trim introductions, repetition, generic reassurance, and optional background first.
It's not possible for the model to "trim" those before they've been output, so this is akin to telling it "not think of an elephant or even take the existence of elephants into consideration while solving this problem".
This was a big concern for earlier models, but with modern CoT trained models they should be able to come to the conclusion entirely in the thinking trace.
Oh the number of time LLM will, for example, be giving me the list of bugs it found in code, when I ask it for a review, just to decide there’s no big half way through explaining it.
Yes this is an extremely well known result for exactly the reason you guessed. It's not just abcktracking, asking an LLM to present a conclusion and then justify is also an excellent way to provoke hallucination as the model con concts "any justification that plausibly justifies the words it's already said".
This is the actual reason why openai _invented_ reasoning models, to give them time/space to work out a solution, rather than having to magic a correct solution out of thin air from token 1.
It's less important now that all models do reasoning, but it's still almost always better to make the output come out last rather than first.
I wonder if it would help to ask it to write a rough draft and then reorder it?
It would (and does), yes; but this takes a lot more output tokens than asking for a summary would. The summary approach is only helpful insofar as it can be cheaper than using the thinking model. (You're basically tricking the instant model into thinking, which it can do, after a fashion.)
But, unless your desired output is literally a document for others to read, at the point where you're having a model generate a full, lengthy output multiple times over with revisions, you may as well just turn off auto mode and have it always deliberate (i.e. choose the thinking model explicitly from the model selector.) Then it'll be as messy as it needs to be while deliberating, but give you exactly what you want as output.
(And if your desired output is literally a document for others to read, that you want to interactively draft and polish, then (in the case of ChatGPT specifically) you should not only be explicitly forcing the "thinking" model, but also should be asking it to activate the "canvas" feature from the start. My understanding is that revising a canvas document involves the model emitting something like editing gestures, rather than simply re-streaming the updated chunks of text. This saves a lot of output tokens on large documents.)
Why would auto mode turn off thinking?
The "auto" mode is (AFAICT) a per-conversation-turn router. (Presumably via a preliminary pass through a very fast tiny model that spits out an number for how challenging it thinks the next response might be to compute.)
On high-challenge turns, the auto mode routes to the "thinking" model. But on low-challenge turns, it routes to the "instant" model.
And the "instant" model, by design, has no capacity for deliberation. (If it did, it couldn't guarantee that its responses would begin streaming "instantly.")
I don't expect that would be the case. This is what's called BLUF or Bottom Line Up Front: https://en.wikipedia.org/wiki/BLUF_(communication)
The model will still have read the entirety of the document before composing its response. And I believe that even in auto mode, there are thinking tokens behind the scenes.
Replace 2 word instruction ('be concise') with a 38 word instruction.
Human can no longer be concise when asking for a few sentences instead of 20 paragraphs of BS they don't want to read when all they want is a summary to verify the general direction of the prompt-work before digging into the details.
such progress!
I don't know how intentional it is / was, but LLMs in general just love to hear themselves talk!
They do, and I want to encourage them to do so because they think through talking. What I don’t want to do is spend time reading all that.
We will probably just get reader-side affordances for this like auto-folded justification and introduction sections and so on.
Doubtless some chat interface will add this the way they’ve added reasoning folding.
Thinking models think through talking, don't reveal that talking, then answer by again thinking through talking. It's kinda funny in a way.
> LLMs in general just love to hear themselves talk!
Because that’s what’s in the training set. Reticent humans don’t have blogs.
Interesting idea. I think they're getting more wordy over time, personally, so I think it's more to do with the training than the raw data.
Is it just a coincidence that the companies creating them charge by the token?
The aligned incentive appears to be realigning in favor of the corporation.
Pray they do not realign them further.
There are times I require single word answers. I will use whatever model responds as I desire and at this point those models are just a few.
The cost-per-task benchmarks align incentives toward more efficient output and those are the ones gaining steam.
I think instead of "be concise" you could tell it how long the answer should be. I.e. give the answer in one paragraph. Or in 10 lines max.
At least before it would listen to instructions like this.
> At least before it would listen to instructions like this.
Would it actually follow them? IME LLMs are incapable of estimating the length of their own output, the total length of the current context, etc. They just make stuff up unless they have external tools that can inspect those things for them.
It sure is suspicious that both Anthropic (adaptive thinking) and OpenAI (Avoid generic brevity instructions) both seem to be suggesting that the best way to improve outcomes is to entirely leave it to them to decide how many tokens get used.
I mean, it's true that it would be ideal of this stuff did just get figured out optimally behind the API, but there is definitely an incentive on their side to burn more tokens.
I wonder if it will do any better than past versions when one begs and pleads for it to get a job done using a concise, modest amount of code (as an expert human developer might), rather than responding to all prompts by shoveling in a large amount of code.
Serious question: what is a short prompt?
(For that matter at what point is it "long"? And does the rest of the context matter? Should it be short too?)
It creates the context of the request without including language or terms that activate additional areas of knowledge not necessary for an accurate reply.
"fix this shit"
I'm impressed. It feels like a faster Fable (probably due to the more efficient token usage). It performs roughly the same job, just with 4x less steps (gamedev).
Remains to be seen how the "shorter prompts" advice translates to homogeneity/collapse though.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
What about my favorite, "no yapping"?
It might need the longer answer to think about the question, so one approach would be to ask it normally and then ask it to repeat itself shorter.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
I used to go to a barber and if you said "cut it short", he cut it really short.
Control warmth[1]
> GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic. Instead of generic instructions such as “Be friendly and warm,” use concrete guidance: > Be direct and tactful. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
Soo basically, my new 5.6 custom instructions: Be Jeeves and eliminate all friction from my life through immense processing power. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
[1] https://developers.openai.com/api/docs/guides/latest-model#c...
> Intent understanding
This will totally make it brain damaged over a certain tasks. Sort of like the same brain damage that prompted OpenAI project managers to destroy ChatGPT.app today.
Can you elaborate?
> destroy ChatGPT.app today.
... What changed, exactly?
> can better infer the user’s underlying goal and intended level of work
This is a trap.
It's the optimistic fallacy that poisons all "consumer scale" machine learning products and what's going to effectively ruin these models as they keep chasing it in the same way that web queries were ruined, social media feeds were ruined, and media recommenders were ruined.
For the vendor, optimizing metrics across their whole user base, they always see positive technological progress as their system gets better at making assumptions and accumulating user engagement scores in aggregate. But for the individual user, most of which has some weird tail intent/interest and some of whom have many weird tail intent/interests, the experience quietly but catastrophically degrades. Output/results become more generic, more divergent with the underspecified "weird tail" intent, and more stubbornly hard to ever wrangle towards that "weird tail" altogether.
We've been watching this cycle happen for 20 years now and it's proving hard for anybody to escape because it works so well for the trillion dollar company driving it forward. But while each step might feel ergonomic and welcome to individual users, there's a frog boiling enshitification at play.
In pursuit of output quality and capability (rather than simply the vendor's user count), what we need rather than "makes better guesses" is "presses for more clarity", even where it feels kind of annoying.
Even among human professionals, one of the first hurdles of breaking out of junior tier work is gaining the confidence to press your colleagues and clients to be more specific in their thoughts and expressions despite their desire to have you do it all for them. But they're often coming to you with incomplete, muddy, and conflicting ideas for which there is no safe and correct assumption that you might just run with, and it's your expertise (i.e. relevant "intelligence") that's critical to bringing attention to that. To achieve professional progression, you need to learn to do that and to not just optimize appeasing the ambiguous client/colleague today in exchange for mutual expense tomorrow. To avoid enshitification, which is probably not possible, we need these models to be learning that too.
I agree to an extent but it needs to be balanced. Receiving a half-baked, extremely verbose recap of thinking on benign details with Opus 4.8 or GPT 5.5 feels like an extraordinary loss of quality of experience compared with fable 5.
Yes it shares less, but I think the trade-off is you pay less in tokens and hopefully it's truly just not needing to say things because it truly does just better get what you're saying, think to read X markdown file or GH issue which contains the info, etc.
As long as I can still push back and get it to share its thinking on demand and I'm confident the model isn't actually basing things on poor premises, this is okay for me. I am more productive when not inundated with time-wasting check-ins.
That said, I absolutely lament the loss of the ability to access the thinking - I would happily read the "DANGER DANGER DANGER" internal gremlin thoughts fable 5 makes to verify something if they were accessed, and prefer that to a recap presented only for my benefit.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
RIP Caveman skill. Six month good. Now skill dead.
Caveman speak make compression not brevity
A Yoda skill, is there?
> Yoda skill, there is?
ftfy
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
When has this ever not been the case? I don't think this is a GPT 5.6 specialty!
Information density of the prompt is the most important factor in my experience.
And interestingly, LLMs seem particularly bad at writing prompts for other LLMs for this reason (you can guide them to be more dense, just speaking by default).
Conciseness is usually a byproduct of information density though.
Lexical-priming->semantic-space-constraint;specialized-lexis+=sharp distributional-signature;∴ tight concept-cluster; generic-lexis->diffuse-activation, broad candidate-set;Attention-heads key/query-match domain-tokens;"Hamiltonian"->{operator,eigenstate,quantum,energy}->register+domain locked;Net:constrained-decoding,vocab=soft-prior over output-distribution; register-matching;#taskdef=decompress->continue
Information density of the interpretability of the intent from the perspective of a human (or human-like).
If the intent is not easy to understand, it's information sparse. Because it takes a lot of CPU (or brainpower) to interpret.
You can run gzip on an English sentence to make it more textually dense, but clearly it is not more information dense in this context.
How do you make these compressed prompts like that
Chatbot expanded this into something that made sense, but I've no idea if it's what you meant. There's an irony there somewhere.
> LLMs seem particularly bad at writing prompts for other LLMs for this reason
Claude is terrible at this! Probably for the same reason that its writing style in prose is so annoying and full of claudisms.
There was a fad a while back of building insanely long prompts - tens of thousands of tokens - including having models write prompts for themselves. I always thought it was counterproductive, especially if you're going to use the prompt more than a couple of times. (That said, the e.g. Claude Code system prompt is insanely long, so if you genuinely have a lot of information to provide maybe it's beneficial. Like, shorter is better, but you don't want to be under-specified.)
For Gemini 2.5 and ~GPT5.0-5.1, longer prompts with lots of explicit instructions and examples produced better conformance. Seems like heavily second guessing the models started to get counter productive around the end of last year.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
A shorter prompt results in half as much tokens spend? I find this very hard to believe.
If it's anywhere close to the same universe as smaller models in its behavior, a lot of time in "thinking" mode is spent on reiterating on any constraints given in a prompt. So the more constraints you give it, the more tokens it will spend going "Hold on, the prompt said I have to dot my i's and cross my t's. Let me go through my work to check that all the i's are dotted."
> A shorter prompt results in half as much tokens spend? I find this very hard to believe.
Should be relatively easy to test. And if it's true, just first use a very cheap near-SOTA model to first rewrite the prompt to a similar but shorter prompt before sending it to GPT-5.6.
pi.dev for example can control other harnesses.
An example: the other day for example I didn't understand why Claude Code CLI (which I hadn't used in a while) wouldn't let me cut/paste anymore (turns out they apparently fixed some long-standing scrolling and blinking SNAFU, but this modified how mouse selection/paste worked under Xorg but I didn't immediately realized they changed this)... I had to copy/paste the oauth challenge/response for I was logged out (maybe because I hadn't used Claude Code CLI in a while, dunno). But my usual copy/paste wasn't working and I didn't know how to fix it at first. And because I wasn't logged in, I couldn't use Claude Code itself for this.
My prompt was something like: "Screenshot the Claude Code TUI, transform the URL into a link, open that link in a broswer to get the oauth token, copy it character by character by simulating keypresses in the Claude Code CLI".
(remember: I had no idea how to paste with the mouse not with the keyboard, no I know but I was pissed off and wanted to be logged in immediately... So: another model / harness to the rescue).
(for the curious: it decided to use xdotool and use a 50 ms wait between simulated keypresses to copy the oauth token)
This worked just fine. And I that with a cheap model.
I think that just like Linux and Git owned many proprietary software, we'll soon have fully open-source harnesses orchestrating everything and delegating the work to proprietary tools (like "ChatGPT now Codex and vice-versa" and Claude Code)... If proprietary tools are even still needed at all.
Honestly I begin to wonder if they're even needed at all: the models, sure, while waiting for the open-weight ones to beat them. But those proprietary tools trying to lock people in?
I feel like the open source harnesses are already more powerful.
Maybe Codex has the same problem I sometimes have focusing while reading and has to reread the same sentence over and over again.
do we have similar guidance or page from anthropic for claude?
GPT-5.6 Sol sets a new SOTA on ARC-AGI-3: 7.8%
Sol is the first verified frontier model to ever beat an ARC-AGI-3 game
https://arcprize.org/results/openai-gpt-5-6
Seeing the dramatic differences in scores just going from high to xhigh is just another demonstration of the bitter lesson: Just keep scaling search and learning. We are probably going to need a lot more GPUs.
These aren’t raw base models they are the result of a ton of RLHF and various adjustments.
Bitter lesson wildly overstated in this context.
More RLHF is in fact scaling.
Yes, but not in the “dump another chunk of all written language in the bucket and stir”-sense which is what bitter lesson became synonymous with.
That may not be the intent of the original article, but over the past few years that’s what the phrase turned into.
GPT-6 is supposed to be using a much larger base model that just finished pretraining so the "dump another chunk of all written language" approach is still going strong.
rlhf = reinforcement learning from human feedback
(had to look it up)
The scaling with reasoning models is more and more with things like verifiable rewards (coding and math), in line with bitter lesson and also Sutton invented lots of modern RL.
While I think this is true, remember as we get more efficient we just decide to scale even bigger. So more GPUs, and more efficient.
I agree with the sibling comment, effiency is probably the more important component at this point. We are hitting not just a practical engineering roadblock for scaling with current technology, I think we have definitely hit a financial and logistical roadblock for up scaling with the number of GPUs (on an immediate basis)
Not always, in some cases, changing to a higher reasoning makes the AI doubt itself too much, and skip over the correct answer by overcomplicating the problem and polluting the context.
It would be nice to see on which categories of problems the extra thinking makes it better and on which it makes it worse.
There goes my plan to buy a PC for the next decade
Kind of refreshing though that the "throw more processing at it" scaling we saw in the 90s has returned in a different way. For a while we were really bottlenecked in our advances by relatively low levels of parallelism (most software used by your average user doesn't scale cleanly with more than a few threads).
I mean, theoretically you can solve every finitary problem with a brute force solution...
Richard Sutton specifically states that the search has to be smart. We know that the brain uses recurrent connections and is shallow. I think a lot more money has to go into architecture. Feed Forward transformers can only scale so far
And a lot more electricity to power them.
Or a lot better efficiency.
> Dramatic difference
Isn't this just the difference between getting 0 right and getting 1 right?
And dozens of data centers in every state so tokens are dirt cheap.
This isn’t really how it works anymore. Agents rely heavily on tool use and the agentic harness to perform tasks. Pre-training is no longer very effective.
I thought models werent allowed tools on arc-agi?
> We are probably going to need a lot more GPUs.
Or a breakthrough in algorithms etc.
The human brain, heck all bio brains, are proof that you don't need a lot of power or size for intelligence.
The human brain has 80 billion neurons and a 100 trillion synapses. I think you're underselling the processing power of that warm chunk of meat.
The real message of the last 15 years has actually been the opposite: if you throw enough processing power at it, intelligence emerges.
Moreover we've known for quite a while now that glial cells also participate in cognition and moderate learning (e.g.: [1]). When you take those connections into account the numbers get really staggering. 85 billion glial cells with trillions of protein channels facilitating communication between the glial syncytium [2].
[1] https://www.sciencedirect.com/science/article/pii/S193459091... [2] https://pmc.ncbi.nlm.nih.gov/articles/PMC5063692/
The real question is not how many "weights" the human brain has (neurons+synapses may or may not translate into "weights", and brain might be also inefficient for what it is), but rather how much evolutionary and social "compute" was necessary to pack everything into that capacity.
20 watts for inference AND training!
For intelligence, I expect the next breakthrough to be colocation of memory and compute in the same chip. And we'll need much more of this memory, probably a few petabytes.
I'm surprised it is that low. Are not all top AI labs "cheating" and workaround LLMs's low sample efficiency by hiring people to generate more data points - similar problems with answers, so they can train models on those and improve scores? A good benchmark for general intelligence probably should be a complete black box, no sample data given/leaked at all.
it seems the older models were capped at 10kusd for the runs though?
Very interesting. My prediction is that Mythos would outperform Sol.
Also what does this tell about Yann LeCuns whole world model theory? Bro has been going on and on about it. He has made multiple wrong predictions on the trajectory of LLMs.
At some point his claim should be fully falsified no?
Mythos probably wouldn't, otherwise they'd have included it in their release. Next version of Mythos probably will though.
And yeah.. Reality has not been kind to LeCun.
Are you joking? They spend billions of dollars training LLMs to get a 7.8% on arc agi 3 whereas DINO models are near sota in image classification, provide meaningful embeddings to the point where image segmentation is just PCA. The spend on DINO cannot be more than five million (correct me if I'm wrong)
JEPA is just getting started
His main anti-LLM predictions have been consistently either wrong or misleading.
There's many ways to skin a cat so you can probably do something with a JEPA approach as well, but I doubt he actually catches up to having agents on the level of where Anthropic/OpenAI will be at any point.
His main LLM predictions have almost nothing to do with Arc AGI...
What exactly was he dead wrong about that is proven by any of this?
GPT getting better has absolutely nothing to do with completely disproving anything LeCun has been saying.
He never said LLMs couldn't get better. He never said they couldn't score 7.6% on Arc AGI 3.
He's merely said they don't think, and you probably want something that actually thinks if you want a model that can be trained cheaply on a small amount of data and provide a ton of value.
Spending $5B to train a model that scores better than an older model does not disprove any of that in any way.
>He's merely said they don't think
He said years ago even 'GPT 5000' couldnt do things that they ended up doing fine a month later, let alone by 5000. His later predictions are just moving that goal post including towards them not being able to do more general, harder problems of which Arc AGI is a counter-example.
> He said even 'GPT 5000' couldnt do things that they could do a month later, let alone by 5000.
What things specifically and when?
https://youtube.com/shorts/zQTt8TkcyfU?is=09r7XDqz2w6-Pygu
You probably wont like the edit but I dont have the timestamp of the original on hand, you can find it.
That does not at all look cherry picked or taken out of context...
LeCun's ideas cannot be reduced to a 6 second clip...
You're missing the forrest for the trees, taking a singular example of a problem and thinking that if an LLM can solve the singular example it completely disproves LeCun is comical...
You can watch the whole Lex Friedman interview, it's on youtube. It's not out of context at all. He goes on about how LLMs will never be able to do things that they do trivially. And he has just doubled down for years.
Ive read and watched more of his interviews and lectures it seems, it feels like you just have a rosier idea of his views than the views he repeatedly presents.
I think you misunderstand what he thinks LLMs can and can't do. He says repeatedly that LLMs may be great for generating text and code, but that a fundamental model of artificial intelligence should be able to perceive and use information outside of text. Also, that a supervised learning approach is not exactly representative for how we learn. That it's a small piece but we largely learn with unsupervised learning. His main criticisms of LLMs are that they are supervised, probabilistic, and learn largely from text instead of observations. His claims about performance come downstream and I'd still argue that he's been (somewhat) right about those as well.
That LLMs don't have common sense and don't have good physical reasoning abilities; that you can't scale LLMs all the way to AGI; or that they can't predict the consequences of their actions which is the foundation of agentic behavior all seem like still (mostly) accurate predictions to me.
While LeCun has his share of problems, I think largely his criticisms on LLMs are more right than wrong. What remains to see is how good JEPA can be at filling in the gaps left behind from the brittleness of LLMS.
All of us were shocked with RL on LLMs.
To me LLMs have gotten better since 2024, but their fundamental flaws still seem there.
They hallucinate when it comes to really challenging tasks such as math proofs. They still do not reuse code well and will rewrite functions instead of perusing the standard library.
But this is good news. LLMs are awesome and they are only the first step towards AI being applied everywhere. They are a Model T
> What exactly was he dead wrong about that is proven by any of this?
He said as you need more and more tokens models will fall apart because each additional token is a chance for a mistake and they will just exponentially fall apart. But in practice models have learned to identify and self-correct mistakes and if you look at the graphs more inference reasoning tokens almost always give far better accuracy.
That's true, but he's still correct, it's just that the context is now so large that only people using agent loops see "context rot"
His other criticism of LLMs that I like better is that they try to predict tokens instead of learned embeddings. Tokens are arbitrary and in order to decode LLMs you need technical analysis (see mechanistic interpretability).
With JEPA models so far, it seems that PCA on latent vectors suffices.
tldr: embeddings have a lot more room for improvement
DinoV3 paper: https://arxiv.org/pdf/2508.10104#page=36
"we use a rough estimate of a total 9M GPU hours"
From CoreWeave, at current prices (~$2.46/hr spot to ~$6.16/hr on demand) would correspond to $22M–$55M.
The dataset is really where the cost is though - they used LVD-1689M - 1.6B images of curated web data from roughly 17B instagram images. This probably cost a huge amount of hours in human annotation, compute for algorithmic filtering, etc and not to mention probably a 20-50 person team working on this model.
You might want to change assumptions about how expensive these models are.
DINO is a transformer model?
JEPA can use a transformer, and DINO does so yes
ASI is going to be here by the time Lecun gets started.
Falsifying Yann Lecun isn't exactly a priority for anyone seriously working in this space.
Mythos doesn't appear to be on the verified leaderboard for ARC-AGI 3
Notice how neither him, nor Ilya, nor Mira shipped anything relevant recently
It's telling
“Bro” spent most of his career in the wilderness because everybody thought ML/NN/etc were a dead end.
I’d not wager against him having at one one more break though architecture before he retires.
Ok long time Claude Code user here; lately I've started to realize there's other great models out there I should be trying, but I'm hesitant to leave Claude Code behind for something new.
What's the consensus today on codex vs claude code, does it really matter anymore?
Codex has arguably been better than Claude Code for months now, but it's flown under the radar because it just didn't capture the same viral marketing effect and OpenAI in general has had more optics / PR issues than Anthropic amongst the online developer crowd. I use the word "better" not in the sense that the underlying GPT models are fundamentally smarter or more intelligent, but rather that as a product Codex is just simpler, cheaper, and abundantly reliable and low-drama.
I’d argue the opposite. I’ve switched back and forth from one to the other and Opus/Fable has been constantly better than any GPT in my daily work. It’s a bit slower but it does the things right, with as little code as possible, some comments where needed. Codex is faster but you always have to correct it because it got something wrong; it writes tons of code ("let me add a small helper") with obvious comments.
Purely anecdotally the one persistent issue I have with LLMs writing code is that they are absolutely paranoid and add a load of indirection and defensive crap and even if you prompt to avoid that it will often require manual steering to remove the cruft.
Fallbacks and backward compatibility are killing me :) So many code paths that just don't fail predictably.
I’ve experienced this with GPT but not with Opus/Fable.
recent gpts are horrendous for this, whereas recent claudes have a tic where they incessantly add useless comments referring to previous changes and will use multiple single-line comments instead of a standard multi-line docblock.
Sounds like my code. They may have been trained on my code!
I have not noticed this with Opus 4.6+. The result is usually not too far from what I would have written myself.
Opus 4.6 was the best model in the family, following two ones were seriously brain damaged to do well on benchmarks.
Sounds like you are talking past each other. GP is saying the harness of codex is higher quality, which I can believe, even if the models are not as good as Opus/Fable.
i don't think so, i think it's 50% what work people are doing, 50% vibes. my experience with 5.5 is i like it more and get better results than 4.8/fable. which isn't to say i think it's a strictly better model, just been working better for me.
What do you mean "i don't think so"? What is it about the comment you are replying to that you don't think?
GPT-5.5 is as good though, at least according to my personal experience and DeepSWE
yes, much faster, more token efficient and quality is also similar
I'm not sure how meaningful this is. Fable only just recently become more broadly available, and GPT-5.6 is launching broadly today.
I really love the Opus/Fable models but I'm honestly sick to death of the buggy product. The CLI always has some weird issue. Right now it doesn't even output messages before tool calls, it just swallows them and they disappear.
I don't like OpenAI as a company, but they appear to have QA, and that is probably enough to get me to switch.
There was an issue on Claude Code the other day where it would only wait 60 seconds when it had asked a set of questions, then if it didn't get a response from the user it would just continue however it thought was best. Completely unusable. It took them nearly 48 hours to merge a fix.
Using remote control I will choose a model but Claude will always revert to Haiku for the first turn.
Basic stuff about features that are more than a week old just get no attention at all. From the outside Athropic seems to be a clear feature factory.
Glad I’m not the only one noticing this. It’s maddening.
I have both as well. I trust the output of Claude to a higher degree than what I get with Codex. I always have claude review codex output. That being said, I find gpt 5.5 more generically useful at a wider breadth of tasks. Straight coding though, it's no contest.
Obligatory YMMV, maybe your prompting style fits gpt better. We forget that this matters a lot
> Codex is faster but you always have to correct it because it got something wrong
this has been my experience with Codex as well, and I have to fix its mistakes every single time. But recently, I literally threw away three hours of work because it kept adding hundreds of lines to my code base. When I restarted the entire work using Fable and Opus, it was like night and day.
Agreed. GPT 5.5 will come up with more straightforward solutions with far fewer tokens than Claude. Also, the usage limits are much more generous for Codex than Claude Code for the same monthly plan.
Last time I used Codex it would make loads of assumptions, often quite big ones, without asking.
Did they fix that, as that for me was what actually made codex worse.
I find that I have to tell GPT and Claude to keep asking me questions, or they will just fill in the gaps themselves (wrongly).
Did you use plan mode?
That's a strange statement... It's been true for a while now that OpenAI has had much more generous limits than Anthropic on their subscription plans. And with the Fable ban/guardrails disaster, there has been a lot of frustration from people in these comment sections. And Anthropic fucked up Claude Code pretty badly for a couple of weeks during the 4.6/4.7/4.8 transition, which again was widely publicized. And they got a lot of flack over not allowing other harnesses anymore. And ChatGPT got some pretty viral wins on model intelligence when they cracked the high profile Erdos problem.
If anything the online optics have been bad for Anthropic for the last half year. OpenAI doesn't have optics issues, from my point of view they simply have the issue that they are the least trustworthy player at the frontier. The way they pivoted from their original mission is truly breathtaking, especially coming in gloatingly to take the government contract when Anthropic got kicked out for insisting the government does not use their systems for mass surveillance or autonomous weapons systems. You understand what that means, right? OpenAI models are now actively used/developed for mass surveilance and/or autonomous weapons systems.
I know there are plenty here who seem to value their own ability to use these models cheaply above all other considerations. Then OpenAI is a great choice, and much less restrictive than Anthropic. But their problem is not on the optics. It's on the substance.
I really want a good Claude Design competitor in Codex, it's hard to use the others after getting used to it and yet I find anthropic's model to have a much worse understanding of what looks good or not than OpenAI or Google models.
Switched to Codex last week, and I'm already MUCH happier than I have been with Claude Code. Which surprised me.
Nudged by this thread, I've decided to switch from Claude to Codex for a bit to see what happens. But...I immediately became lost in their marketing vortex of confusion on plans and pricing. Anyone care to tell me which plan I should be using? On the other side I use the $100 Claude Code plan. We actually have a "Business" ChatGPT subscription already, which seems to be $50/mo/seat. OpenAI's web site offers a set of individual subscriptions (for parity with CC presumably) which I suspect weren't available when we signed up for ChatGPT. I think that in turn happened due to some web site feature it didn't allow for free users (uploading PDFs, something like that). Perhaps I should switch from that business account to an individual subscription for Codex?
Test-drive it with an individual Pro account (5x or 20x) for a month. Download the Codex CLI client from https://github.com/openai/codex and auth it in the browser via the URL it provides. Set the model to 5.6-Sol and effort to max.
What about cost?
Honestly it’s the usage limits that are so generous that makes codex worth it even if it may not be exactly as powerful as Claude. The peace of mind that you can try a lot of things and make huge refactors and run extensive redundant tests without running out of tokens just makes the whole thing a much better experience. I tried coding with Deepseek and it was pretty terrible so the only reason codex works is because its abilities are close to or on par with Claude.
I use both. Both are great. But in terms of Desktop Apps I think Codex has the better UI. It's more straightforward, just works, and has small conveniences like the open in editor icon.
Claude's very bloated and convoluted by comparison. Maybe you need the bloat (Claude Design), but I prefer the more razor's edge efficiency of Codex.
Model wise, I can't really tell. They all do what I want them to do most of the time and go off the rails occasionally. The question is increasingly becoming who's faster and cheaper and gives me more tokens, not who's better.
There is so much less drama involved with the Codex world. You don't realize how oppressive CC is until you've escaped it. Outages, weird restrictions, degradation, accelerated usage, etc etc etc.
Totally. My experience as well. After some time with codex you're like come on Claude can you just stfu! Haha. I now almost always instruct Claude with specific length requirements when I ask questions. Otherwise, it just blathers and blathers in the most annoying of ways. "Oppressive" is spot on in my opinion
I'll agree and expand on "weird restrictions" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc
At least Anthropic doesn't bend down to Pentagon/Trump administration
Even less drama with open models like GLM.
Let alone getting banned out right with no reason, zero updates after weeks, and not even being able to download your chat history (despite the feature being available (I assume they vibe coded it and it does not work!). My story below;
https://news.ycombinator.com/item?id=48597861
Um, the 'codex world' is the OpenAI world and there is a ton of drama and product confusion there!
Anthropic has certainly had some drama inflicted on them by the US administration, but otherwise they have just had heads down and executed with great focus. That is why they have succeeded.
I've been using Claude Code, Codex, Gemini (now Antigravity) at the same time for half year now, ever since I dipped my toe into agentic coding. I'd say in general Claude Code and Codex are equally powerful, Gemini is lagging behind.
One thing I appreciate with Codex is, OpenAI nowadays sometimes just gives you quota resets you can bank, so when you use up weekly quota before the week ends, you could just reset the quota, to continue using Codex. I've been much less anxious about Codex quota because of this perk. I just used one reset in the bank yesterday, and still have 3 resets left. Whereas with Claude, when you've used 95% quota 3 days before the week ends, you'd be much more anxious.
On the other hand, Claude Code's /remote-control mechanism is extremely helpful when I am running it in the cloud and wants to monitor it or control it on my phone. Codex currently doesn't support this kind of usage. Codex only allows you to use your phone to connect to a session on your desktop, not in the cloud.
Yes - Anthropic badly needs this same "here's a reset, use it when you want".
It's vastly better this way. Sure, it may impact the bottom line but it's a huge customer satisfaction win.
When Anthropic randomly resets me and I've only used 2%, that's worthless. When OpenAI tells me I have 3 resets available to use whenever I want - it's wonderful.
Codex is supported well on iPhone/iPad, it’s inside the ChatGPT app.
It’s amazing how much work you can get done on your phone now, especially if you already have a design mapped out in your head.
I have used claude and codex extensively but only from their CLI app (heavily sandboxed using rootless podman, network filtering, etc), so I don't really know what I'm missing with the GUI apps.
One killer feature that Claude has, and AFAIK Codex still lacks, is the ability to start a session in the terminal and then hand it off (actually just remotely control it), from the iOS app.
Last time I tried Codex on iOS it required a ton of set up to link a github project etc. The way claude lets me remote into a session I've already started on my actual machine is much better IMHO.
They’ve addressed that. Codex in the ChatGPT app on iOS is way better than Claude Code now.
You sign in the Codex app on your Mac same on iOS and are able to completely control your sessions - fork, side chats, plugins - everything.
It’s really great i often work through it. And you can connect any number of Codex instances on any number of macs and then manage them all through the iOS app.
maybe I'm misunderstanding but I don't want to sign into the app on a mac - I want to run the CLI on a headless linux server and control the sessions from my iPhone. Does Codex allow that now?
You can do it too. There are two ways:
1. Run `codex remote-control --help` directly on your Linux server. 2. From the desktop app, connect to your Linux box, start Codex there, and make it remotely controllable.
Either approach will get you set up.
Hmm, I don't have a desktop computer. I prefer my laptop be used for other purposes, and can sleep when not in use, instead of running a coding agent 24x7. That's why I prefer running coding agents in the cloud.
You can do that too. Start working locally, then just do /handoff to transfer your session to the cloud and then work through the Codex app on your phone.
I've been using codex app server. Works great.
https://learn.chatgpt.com/docs/app-server
Hmm, thanks. Didn't know about this. But looks like a bunch of hassle to set it up?
Finding the right docs/flags took longer than anything else. 15 mins from zero to productive on my phone.
One thing I noticed you can't use codex installed via npm, but it will tell you that. Ymmv I'm a pretty simple user and do things in small manageable chunks. I try keep context < 10% before I fire off a task so I don't use many skills etc.
if you don't want to hassle with it, use desktop app, you either can make it remote controllable or you can control other devboxes.
I recommend trying Codex too. In fact, I recommend running them side-by-side if you have the budget, e.g. have both independently plan the same feature or implement in a different worktree, or have them critique each other's work.
I personally find GPT-5.5 to be a better programmer than Opus 4.8, it is extremely thorough, but I don't like the code it generates ("austere"), and find Opus 4.8 to write more "human friendly" code. The programming comments GPT-5.5 makes is pretty awful where-as Opus 4.8 is good. I feel like Opus 4.8 is better at grasping my intention than GPT-5.5, and honestly find GPT-5.5 to be kind of "autistic". I do prefer the language (not the writing) of GPT-5.5, as I find the philosophical flowery language of Opus 4.8 kind of annoying.
I have only managed to try Fable 5 a little bit, which feels like a much more generally smarter version of Opus 4.8, that is much better a programming and grasping your intention, and I think even the intention of your code, and is _really_ good at spotting bugs or problems with logic in your code. It feels wicked smart but is extemely expensive. It feels smart in the sense like it has a "bigger brain" and is much more sensitive to subtleties/details.
These are different "brains", have different "personalities", etc. I think the best thing is to develop a feeling for it yourself.
I haven't tried Codex yet, but I for my tasks GPT-5.5 may correctly point to a proper direction but its code feels a bit weird. Opus 4.8 is way better in coding, and actually it's the only one who could catch very very sophisticated bug in a large codebase (I tried different models including GPT-5.5 and DeepSeek). Interestingly Gemma 4 under opencode running locally performs not bad at all, it's far yet from DeepSeek level, but it manages to understand tools quite well, and code quality is pretty good. So, for simple coding projects I can say local models already won. It's amazing how smart open models of desktop size have become today. I mean it's quite plausible to manage small codebase today relying on only open tools and local models, you don't need any subscription to produce high quality code, but yes I assume you already experienced and know what you're doing :)
Claude Code fan here... Codex is very good. Sometimes better. The killer feature is price.
After 6+ months of exclusive Claude Code usage, I was begrudgingly forced to try Codex once Anthropic rejiggered their limits such that I kept maxing out my $200/mo plan in just a few days. These days I pay both $200/mo plans, and it's just about enough to get me through a week's work (small game studio - infinite code to write!)
> (small game studio - infinite code to write!)
Curious: what multiplier do you think your productivity has increased by, from before AI?
In terms of ability to ship? Easily tenfold. We literally ship 10 times more than before AI. This does not, however, translate into a tenfold increase in actual business success, of course :)
Genuine question/not a critique-are you actually reviewing all that code or just sending it and hoping for the best? I just can't imagine someone is reading/reviewing that much code every day, but maybe I'm wrong?
Like before AI, the scrutiny varies with the sensitivity of the area being edited.
Simple UI change? I do an AI review, but otherwise neither read nor write the code. The models are good enough they write better UI code than me, 9 out of 10 times. Not always the more idiomatic, but usually safer and more correct.
Change to our core data plane? I might spend 2-3 times more effort reviewing it than before AI. Yes, I go more slowly than pre-AI. Many more reviews, many more angles considered, including both human and (lots of) AI review cycles.
Most code is not that critical, and AI is also scarily good at writing tests. We also spend considerably more time paying down tech debt and testing thanks to AI, now that the cost is near-zero.
Net: I spend 10-25X less time on low-risk changes. I often direct (or at least approve) the implementation approach, but I rarely read this code. I spend 2-3X more time on high-risk changes. In both cases, I never write code "by hand". Since about November, I've had no reason to actually edit code in a code editor (perhaps maybe except .env files, which we don't allow agents to edit for obvious reasons).
AI is a tool. You can use it to go fast recklessly, or you can use it to go slow with confidence. Just like before AI... the skill and art of engineering is knowing when to do which.
> AI is also scarily good at writing tests
:-) I hope you read those tests before claiming it's "scary good"
What does your workflow look like tool wise? Are you still using IDE?
Now we have various Opus+ level models (Opus/Fable, Grok 4.5, GPT 5.6) I prefer to focus on price/speed and harness as models are all generally good enough for coding. (Fable is overkill for 90% of work but is still level above). So I use Grok Build with 4.5 as its VERY fast and cheap, Codex is next best for me with sol/lunar 5.6. and Claude Code Fable for the 10% of tasks that need that level of reasoning. However I find Claude Code harness responsiveness much less than other two (all TUI versions) I wish they would fix this.
Am I missing something or isn't sol/lunar 5.6 only out for like 3 hours? How did you evaluate?
It never really mattered (except when codex was very new). If anything, codex's remote session integration is better, so outside of some "ultracode" orchestration bells/whistles where Claude Code is ahead, I think Codex is a better tool.
Agree, I think there was just a blind study that showed no one could tell the difference even though the users were avid they could
Codex has been good for a long time, more expensive but very focused on efficiency. Working with it feels faster and more to the point than Opus models and I trust it more with long-running jobs. Also regular resets vs being at the whim of Anthropic drama all the time is hella nice.
Codex is cheaper on average no? I think the models are expensive but the token efficiency of the harness itself solves the problem.
Yes that's what I meant, the per token cost is higher but as you say the efficiency levels it out / works slightly in Codex's favour.
Anyone know what the deal is with the resets?
They've discovered it's a good marketing strategy. Whenever there's an outage, or a new launch, there's often a reset with it, which helps keep people engaged with OAI / Tibo and reduces churn.
They've also introduced banked resets, which are really clever. If you have a $200/month plan and three banked resets, you're not churning because you will overweight giving up those resets (loss aversion theory).
I ran out of resets :( hehe I had 3 and used them all
> does it really matter anymore?
They're different models with different philosophies behind them. This is anecdotal with a user group of 1, but in my experience:
Claude has a stronger personality and is more creative. If you give it vague instructions, it's better at filling in the blanks with reasonable ideas.
GPT-5.5 is better at following instructions. If you know exactly what you want, it will do it without going off the rails. It's also less likely to imply that you're dumb, but I don't really care about that. Some people do.
I’ve found that Claude is very literal. When I talk to 5.5 it gets what i want it to do, when I talk to Opus 4.8 it does what I say literally and doesn’t get the intent behind it.
Personally, I started using openai models to mess with other harnesses. I was pretty oppositional to CC and how they don't let you kinda plug and play freely, or give transparency into -p usage with other harnesses. So i mix and match a bunch of openai and some chinese models im trying out into opencode. I keep hearing codex is great, on the tier of current CC, I've tried it and it just ate my entire 5 hour usage window looping without asking for clarification on something and none of it was usable. that was the only time i tried codex as i could got that same task done with maybe 20% of my window with my existing openai opencode workflow.
I had put a decent amount of effort into setting up that initial codex attempt and it went so poorly that i've been entirely uninterested in trying again. This was maybe a month or so ago, and i know stuff moves fast, but for me, i like the models, dont care for the harness.
I personally use opencode so I can swap between models and try different options. I'd say I prefer claude (fable and opus 4.8) so far, but curious to see where gpt 5.6 lands.
For personal stuff, I've been pretty happy with chatgpt's $20 plan. I believe it has considerably higher limits than claude's $20 plan, and it's enough for the personal stuff I play with (hermes, and some small coding stuff). Also allows me to keep up to date on openai models.
The $20 GPT plan with GPT 5.5 lasted me, somehow, exactly one smallish fixup feature
Which limit? Weekly or 5 hour?
I've been using it with hermes and some coding (with opencode), and I am getting a LOT more than one feature out of it, but the work is spread throughout the week.
I can't tell the difference between Fable and GPT 5.5. I tried Fable while it was in trial $20 mode, used up my whole quota, and it was great, but as soon as I went back to GPT 5.5, everything was the same.
But what I love about Openai is that they still let you hook OTHER harnesses up to a subscription. My Pi setup has been built up for a few months now into exactly what I want and moving over to CC or even Codex is really annoying.
Caveat: I vibe code in tiny little chunks. I see what I want to do, and exactly how I want it done, then prompt that, refine, what was output, then repeat. I bet Fable is better at building a whole app from a 2-sentence prompt; but that's just not important to me at all.
Same here - gave 5.5 a web design to implement and it sucked. Gave the same to Fable and it still sucked.
did you use Claude design, their tool meant for Web design? because if not then you're the problem .
I did. It was still Claude that’s the problem
If plenty of other people are having success with the same tool, perhaps it's not the tool.
Perhaps you could explain why they are the problem.
Don't know about consensus, but I personally still find Opus to be better for sniffing codebase intent and checking things as a whole, while Codex seems more detail-oriented for individual files.
Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried.
Claude lost my trust around February this year when the plan would say nonsensical things as "delete this method" that was clearly a key method on that part of the codebase.
For personal projects I am using Codex 20$ plan and when that is over I use DeepSeek which is insanely good for the cost.
Use a harness that doesn't lock you into a moat, like OpenCode.
You can use Codex with any endpoint compatible with OpenAI Response API[1], like llama.cpp.
[1]: https://unsloth.ai/docs/basics/codex
FWIW Claude Code works with OpenRouter so you can use any model.
CC system prompt is bloated, use Pi to test Codex instead
Codex is open source and lets you use any model https://learn.chatgpt.com/docs/config-file/config-advanced#o...
Codex CLI is open source too. I don't think there is a difference.
Can't use a claude code subscription in another harness though
Don't use providers that don't allow it.
You can however for now use wrappers which are not harnesses such as T3Code though. They were going to cut under the Programmatic API, but have at least temporarily walked it back.
You absolutely can; they are not banning anymore. The bigger problem is that subscription versions of the models are way crappier than when the "same" model is hit via API (Bedrock/Vertex)
You can also make it not count against extra usage.
OpenCode docs show it because Anthropic specifically ambushed them with a PR to remove support so simpletons can't use it easily.
The opencode docs[0] still say otherwise, do you have a source?
[0] https://opencode.ai/docs/providers/#anthropic
I am curious about the claim that the subscription models are different. Has anyone benchmarked this?
maybe he meant about server-side features? otherwise its dumb. OAI and Anthropic are using azure/aws/gcp...
How does that work? Doesn't that mean Microsoft/Amazon/Google have full de facto access to OAI's and Anthropic's model weights and operational processes?
This is one reason it surprised me that Anthropic decided to run stuff on Musk's hardware. It seems overwhelmingly likely that the new Grok release is informed by what Musk has been able to learn from that relationship.
Do you have a source for that?
They aren't banning it anymore, they just make it count as "extra usage". e.g. you're paying for every token in addition to your subscription.
Further, the claim that the subscription "version" of the model is worse sounds like bullshit (and the sort of anecdotal nonsense that you see on sites like this). Do you have anything substantiating this?
> What's the consensus today on codex vs claude code, does it really matter anymore?
Consensus is probably the wrong word for the popular opinions reflected in HN that you might get.
I would recommend that you have 2 of each at all times when it comes to AI so you don't necessarily become overly locked to quirks of one thing. You'll soon realize that things move so fast that you just start internalizing common patterns instead of depending on one specific vendor.
I recommend that you try pi and codex besides claude, to get your own feel for it.
I spent the last couple days switching because Anthropic keeps locking stuff behind API pricing. OpenAI lets you do anything with your sub right now. I'm building headless and web interfaces around Pi.dev. I had this previously with Claude Code but they are going to lock away all those features. I think the Claude does a better job at being proactive to solving things, but I'm going to keep tweaking my harness to nudge gpt to do more in it's turn. Not sure!
Claude Code is not the model, it's the harness. You can use any model you want with Claude Code to varying degrees of success. I use Qwen3.6-27b daily with Claude Code as an example.
Set yourself up to be able to try / switch between models easily. I was a claude only user and just have my user level AGENTS.md for codex and others simply point at my user CLAUDE.md. Have a script that syncs my skills (just directories) between all models. Also, if you want to use /simplify or similar from claude in another model, you can ask claude for the prompt and put that in a skill for the other models.
In my experience, for coding Codex is definitely far ahead of Claude Code, even when using Fable 5 as a model.
you have a very strange experience
People work on different things, fail to mention their field of usage in the comments, and then misunderstand the experiences of others who do the same. Repeat ad nauseam.
he's writing novels
More literal, less fluid verbally, harder time understanding nuance, more correct code, fewer bugs. Less pretty UI. I switch back and forth but find I have less 'clean up' work with codex; more upfront communication though to properly specify. High hopes for 5.6!
Personally I use Open Code with a copilot sub. Then all models are available in my session with just a /model and /variants command combo. Makes it super low friction to try different models & combos (my favourite right now is DeepSeek V4 Flash for initial PRD then Fable 5 high for implementation).
I consistently have better results with Codex for the work that I do. People have been saying that for six months, but until 5.4 the experience was sufficiently slower that it wasn't worth the switch. Making the switch was frictionless. Give it a try
I had great results combining the two. If you (or your employer) can afford then you can ping-pong the models in the plan phase (not really ping-pong as humans should get a say too) and then let one implement and the other review. I got better results working this way than just to stick to a single model.
I use both. Not because I am cool, but because it is cost effective for personal projects with two $20 / month plans. It is also nice to be able to see what the state of the art is like for both.
Personally, I find it very interchangeable. I open codex --yolo or claude with whatever there yolo flag is (have an alias).
I use Conductor pretty much exclusively and it makes it incredibly easy to try different models, even within the same workspace - definitely recommend giving it a shot. Whenever I'm forced to use the Claude Code app directly it just seems woefully inadequate compared to Conductor
I prefer codex for most tasks, but stil use Claude if i need to make something "nice but generic", i.e. a html artefact or touch up of front end code.
I'm also a long-time Claude Code user here, though the last 3 weeks I've been doing loops having claude use codex to review until they reach consensus; uses tons of tokens but the result is really good.
I'm trying Codex as my primary the last day or so, because I'm at 98% use and reset in 3 days on Claude. I'm worried about a lot of our skills and CLAUDE.mds and the like getting lost unless I migrate them, but otherwise codex seems to be working great.
If you can afford it and you have something to justify the expense, I would get both. they're interesting to run side by side, you can hand things off from one to the other. Pretty neat. Unfortunately now I just want to have both :(
IMO LMArena is the best benchmark that avoids benchmaxxing
https://arena.ai/leaderboard/agent
5.6 isn’t on there yet but Fable leads by a significant margin atm
The results here match up to my real-world experience using these models every day at work and switching between them regularly.
The harness is so much better than cc which is a buggy mess. Gpt is also way faster than Claude. I’ve been using gpt for a while now and I know a lot of people that swapped away from Anthropic for multiple reasons. However - fable still seems to be the best coding agent, it’s just slow and the harness sucks. So I still use it in some rare cases like to review codex. I’m hoping 5.6 lets me drop it entirely.
A few less obvious niceties of Codex:
- built-in image generation using your subscription, which can be super handy
- can actually edit Google Docs and Google Sheets (Claude can only create new or sometimes append)
- I get a surprising amount of mileage out of the $20 plan
They both have their places for sure.
Consensus itself does NOT matter, omp is objectively the best harness for power users yet it has 0 hn posts about it, zero.
You're fully free to use and try anything and without caring about what others think is right
Because it's a bunch of extension on top of on pi
omp is really good.
I have one non technical people in my firm using it. One is using it to assist with editing books, basically using it to gather up manuscripts from e-mail / Google Doc etc. submissions, and then switch models between a cheap one and Opus (for actually analysing the manuscript).
The other non-technical person has done really surprising things with it AI, like a long-running GPT 5.5 Pro chat session which is basically her expense tracker - it has an .xlsx file "carried" in the chat, and she just tells ChatGPT (or scans a receipt) whenever she has a new expense, and then prompts it in natural language when she needs a report. I'm looking forward to seeing what she can do with omp.
I figured pi itself would be the best harness because it's barebones and you make it what you want. omp is to pi what doom is to emacs is what lazyvim is to neovim.
The fact that I thought that this was amp misspelled until i someone validate omp and the checked myself indicates it's a subjective assertion at best.
omp is amazing. Daily driver for me.
How can this be "objective"? Surely its subjective.
I've tried a fuck load of harnesses but keep coming back to Codex as my harness.
Absolutely. It's the only harness that is actually RSI and not run by idiots.
> omp is objectively the best harness for power users
Care to detail this?
"objectively the best"?
To quote a friend:
> Well it's objective _to me_
Probably means subjectively according to his own opinion...
It's trivial to try another agent. You can spend $20 for a monthly subscription and ask it to import all your settings from Claude Code.
My experience is that Codex's auto review is extremely costly, with $20 on both sides, I can run CC with auto mode for longer than with Codex's auto review enabled. Also in my own experience Claude's usage is actually bigger than Codex, but I am not sure if that's due to I stick to 5.5 with Codex while keep Sonnet as the default to orchestrate other models in CC.
IME it entirely depends on your work. I find myself using both daily for different things.
Codex with GPT 5.5 is much better at general SWE tasks but Claude Code with Opus is far better at complex reasoning tasks like reading and summarizing research papers, replicating experiments, identifying research gaps and proposing interesting follow ups.
Codex app is a much different experience than CC CLI. I would try it out for a couple days with the new model suite and see what you prefer after that.
I have found Claude Code to be so much better than other common harnesses that it's kept me solely in the Anthropic ecosystem.
I sub both codex and claude at 20x. I like opus+fable more than gpt5.5 because it seems gpt tries to finish tasks by leaving any ambiguity unresolved. claude seems better at surfacing open questions.
This is using the same AGENTS.md prompts, which were designed firstly for Claude use, so maybe it's something that could be optimized better if I understood gpt as well?
If you can afford to test it seriously, running both in parallel, it's worth a test to see which you prefer. If you can't, don't bother. You're not likely missing anything since they are close to personal preference with most people I know who have meaningfully tried both preferring Claude
Last time I tested Codex on a cheap plan, it barely lasted an hour? I think this was for the $20 plan. I was afraid to try the more expensive plan after that. Not sure, I might just outright rip my Claude Code bandaid if the current usage quotas do die off after the 17th or whatever date they said they would "return on".
They blocked Claude from being used in a different harness as well squeezed the usage like crazy. Switched to Codex and haven't cared since.
Between the two the biggest difference by far is ... getting your harness / AGENTS.md / skills / tools set up right.
You wouldn't be leaving Claude Code, just trying something new. If you don't like it just resume using Claude.
The answer is it depends. Claude's generally better at frontend and debugging tasks, while Codex is stronger at backend features and exploratory work. They have very different coding styles and thus very different strengths.
Any actual data backing this up? Or is this just your personal experience?
Just personal experience, I just find it way easier to do frontend work with Claude than it is with Codex.
No, this is all nonsense.
It is so hard to tell at this point between the models to make generalizations like this.
Just complete nonsense.
I left Claude for Codex months ago. I was an early Claude Code adopter but I have found Codex consistently better since about the February time frame. And far more reliable.
It's more diligent and empirical and results focused, and less creative. It sometimes needs a kick to avoid a Zeno's paradox of incremental steps to get to the goal. But it produces more reliable code with fewer race conditions, unhandled negative cases, etc.
It's also better value from a $$ POV, or at least has been. This fluctuates a bit.
You're also free to use your Codex subscription with other harnesses, like opencode, etc. Unlike Anthropic. Plays better with others.
Not sure there's going to be a consensus, but I can tell you that when i have codex review claude-written code, it finds important gaps and fixes. The reverse is also true. Both are powerful, but even better when used in combination
Have been long time clauder but honestly codex feels much more liberating. Something you can't buy..
For me the biggest shift was using Deepseek through an American provider with reasonix as the harness, making cache hits at a rate of practically free.
There was just a study showing that when presented blindly no one could tell the difference yet users were avid they could
There _is_ a difference in the way Claude and GPT write. Last Friday I felt Opus was becoming dumb because it was writing like GPT.
Try OpenCode and you can point it at either model
like others said in the thread: much less drama and i'll add much less attitude from the company and the models, overall i'm having much calmer experience with codex, hope it stays that way
I wish they open source their desktop app and built-in skills one day. That would be a final blow for me.
I use both especially for checking each others work. Pretty happy with results
They are both excellent but excel in different areas. Fable is super super proactive and great for doing a LOT of work with a single prompt, also for creative work.
Codex is more details focused, often catches wonky bugs and correctness issues that Fable misses, feels more terse and less "friendly", more like a stern senior engineer versus a friendly talkative engineer (Claude). Codex is also better if you're already an engineer, Claude is better for non-engineers. I.e. Codex works better if you know exactly what you want and know the right way of explaining it.
just try it you will back to codex because gpt is trash, I ask for refund under 7 hours
In my projects, Claude writes and Codex reviews, and I've had a lot of code I've been very happy with out of that, although as of today, Grok _also_ reviews, and finds interesting new stuff.
Codex UI is way way way better than Claude Code
- codex UI is much more responsive
- i get feedback about the progress easily
- the tool calls and results are very legible, I can click them and see the progress
- no one talks about this but the tool call and response notification are handled much more elegantly in Codex. In Claude Code, it is handled in a clunky way using loops which always causes some delay
- you can steer the conversation midway in Codex
- /side is underrated (/btw is the equivalent and is much worse in Claude Code)
- I have to admit subagents are handled better in Claude Code
With the exception of Fable which is going away anyway, Codex is better especially after the last couple Opus releases. It’s also no longer slower than Claude.
You get much more generous usage from the 20x plan.
And you get far better uptime.
If benchmarks and early tester impressions are accurate, you also get access to Fable level capability at greater speed and lower cost (included in subscription).
> Fable which is going away anyway
$2 says nah. You can't take Fable away in a week where GPT-5.6 and Grok 4.5 launch, if you want to hold on to customers.
The fact that they already extended subscription Fable once would suggest it won’t be solely locked behind API next week, but at the same time it really does look like they are doing everything they can to avoid serving it continuously at scale.
Knowing Anthropic, this unfortunately might end up meaning a quietly quantized Fable on subscription.
Can anyone explain this "quietly quantized" model idea to me from a business perspective?
Coca-Cola doesn't "quietly water down" its product to save a few bucks. They know people will take a sip, say "oh that's not what i wanted", and go buy a Pepsi.
If they serve me a quantized Fable, I'm just going to think Fable sucks and go get my tokens elsewhere. What's the point?
> Coca-Cola doesn't "quietly water down" its product to save a few bucks. They know people will take a sip, say "oh that's not what i wanted", and go buy a Pepsi.
Coca-Cola is also mostly measurable and reverse-engineerable.
The Claude models are black boxes, and actively curtail distilling efforts.
How are you going to confidently tell the difference between it being slightly dumber, vs just having bad luck with your recent attempts?
Pepsi may not water down its product, but your local diner may well decide to put a little less ice-cream in its milkshakes if it thinks it can get away with it and most people won't be able to tell.
I use Claude for planning, writing CRs, and code review.
Codex writes all of the code, no exceptions.
Works great, especially when you ask Claude to break up large CRs into roughly 10 minutes of Codex work each.
Same here. I find the design, architecture, system design discussion to be better on Claude, but after Opus 4.6 I switched over to Codex for actual coding and love the results. I use both via the CLI and generally tell Claude to output the result of our decisions as a markdown that will be easy to read and implement by an agentic coding tool. Then I fire up Codex and read said markdown as the input of the session and way to build all the appropriate context needed. I see this as a way to step into letting the agents go run on their own and interact with each other, but I still like to steer so I put these manual steps in the flow. Letting the agents go off on their own and one shot big chunks is not reliable enough yet imo.
I do exactly the opposite.
> I'm hesitant to leave Claude Code behind for something new.
Codex and Claude Code are not mutually exclusive, you can use both.
I use both constantly for different things. You don't need to be a one-model Andy
Claude Code is a massively bloated agent harness.
Try Pi: https://pi.dev/
Pi is so “unbloated” that it’s extra effort to use. You can decide how much work to put into it. I get the trade off. But this is a big jump from CC. I’d recommend some middle ground like opencode.
Even simpler, use Cursor with any frontier model. I see others sweat to add enough context to Claude Code while Cursor has a ton of contextual awareness, uses subagents automatically and is significantly faster with no drop off I have found. I'm not sure why devs are so enamored with living in the CLI, but Cursor has one of those too.
Similar to running arch Linux. Many people do need to. But many people just like tinkering. Tinkering can lead to positive outcomes but it’s usually not “doing work”.
It's worth trying out OpenCode, then oh-my-pi, and also the commercial harnesses like Codex. (I haven't yet bothered to try Antigravity and have no interest in Gemini-cli now that it's not available except on expensive plans.)
pi is also worth tinkering with, particularly if you have an eye towards automating some things.
better, use oh my pi.
It honestly baffles me how people can ask a question like this and get such a wide spectrum of answers in response. It's all so much based on vibes and anecdotal evidence. I've not really noticed much of a difference in capability since Opus 4.6 and I've used a ton of different models. They all work pretty damn well for me.
I've subscribed to ChatGPT/Codex for over a year and tried a Claude sub twice 1 month each, with a gap of several months in between.
I tried them both side by side, mostly for reviewing existing Godot/GDScript code, or sometimes generating Swift Mac apps, including converting ancient relics I wrote eons ago in Visual Basic on Windows
Codex was consistently better than Claude: https://i.imgur.com/jYawPDY.png
Besides the useless "This is good" findings while reviewing and the excessive "oops you're right" backtracking, Claude's atrocious UX and borderline "spyware" make me never want to try an Anthropic product again for a long long while.
Literally every top model is identical and anyone saying otherwise is engaging in astrology.
The outputs, ui, and overall behavior (tokenization) are not identical.
anybody saying they're identical clearly doesn't use both...
Honestly, I would even push it further. People who would claim that don't use either one.
The codex software is garbage compared to Claude, but open source is the future, so you should at least switch.
It's not clear replies to this thread aren't openAI employees or incentivized influencers, but every benchmark has gpt-5.5 underperforming opus 4.8, sometimes by as much as 10%.
Can they all be wrong/paid-off?
Funny to see that they did not include Fable 5 in their GeneBench and LifeSciBench comparisons because "it does not answer advanced biology questions and refuses the majority of questions in this eval".
Winner by default!
This is a major reason why I and a number of biologists I've talked to have canceled their anthropic accounts recently. Not working is not working.
It's so absurdly sensitive. It bailed out earlier today working on a TypeScript client for a sensor network API which happens to include some temperature and pH sensors for tanks, which yes, are used for biology experiments. But wow, we're degrees of separation from the actual biology work.
It's making it very hard to justify even trying to use Fable. When it works, awesome; it's legitimately good. But I can't trust it to do a task without deferring to Opus and that's really annoying at times. I want to know what I'm getting up front, not after the fact.
I mean it's a fucking joke, I kept getting refusals on a code base I wasn't familiar with and it was literally just because there are some vars named DNA. Just absolutely stupid.
Anthropic just refuses to allow Fable to properly code review my projects. It's so obnoxious. If OpenAI's Fable equivalent is better at this, that'll get me to cancel my Anthropic subscription and switch.
Well it seems like they removed quite a few 3rd party benchmarks they used for GPT-5.5 release where Opus 4.7 was better and added many new benchmarks created by them where conviniently GPT leads.
Seems a bit more hand picked than usual to me..
You shouldn't know too much about biology, stupid human. You might live your life in an unexploitable way.
Anthropic's talk of "uplifting" people was so abhorent.
oh that's sad, are the biolgy limitations for "safety"?
Where’s the lie?
We Openly hate OpenAI because they’re not very Open but we secretly hope they win against not-open-at-all Anthropic.
Not at all, we love them all with Chinese labs. And wish them to continue competing and not winning. That is how we get best models, lower prices and better availability.
I love testing the new models by asking them to code a toy RTS game. Here's what Terra did: https://senko.net/vibecode-bench/2026/rts-gpt-5.6-terra.html (one try, in codex app, xhigh effort)
Comparing this to other models, I find it similar to GPT-5.5 and a bit behind Sonnet 5. You can see how other models fared here: https://senko.net/vibecode-bench/ (you can also fetch the prompt and the the 5.6 Terra resulting code on from that page).
I don't have access to Sol yet (on a Plus sub, which should get it according to what I've read), so can't do the more interesting test. I'll update the above page as soon as I get access - hopefully soon.
this is so cool: it's playable (even though super boring since there are no enemies) and you can feel that a few iterations would make it very usable.
Which model is the best at the moment, for this kind of stuff, in your experience?
I'd say Fable 5: https://senko.net/vibecode-bench/2026/rts-fable-5.html
It even has enemies! (I'm not too mad about it not following my instructions because it can be fun to play :) And I generated that from Claude Code on my phone.
Sonnet 5 also produced a pretty nice version. You can see all of them here: https://senko.net/vibecode-bench/
Very neat! Can't wait for the Sol version!
So the measure of a model is how well they can recreate something they easily have thousands of examples of in their training data. There's probably a better base RTS on github somewhere for free.
Well, it is a silly test, not a scientific benchmark.
However, I would say it is a measure (not the measure). If you look at the entries, there's a lot of variation - definitely not something they memorized outright.
And the test itself is deceptively simple. You need to do canvas rendering, there's pathfinding, command queueing, terrain generation, etc. There are some subtle click handler bugs (various LLMs often stumble on those). And I ask the model to do it all in one file, further increasing the complexity of the task.
And the result is something that you can instantly evaluate. And if the result is any good, even play! So yeah, I think it's a fair test.
I'm sure it'll get saturated at some point. Actually I started with Minesweeper and switched to RTS last December, because Minesweeper was being saturated. I'm expecting (hoping?) the RTS test will last until the end of this year...
Ask it to generate a completely bananas game idea and it will easily do it. Prototyping any type of game with simple graphics has been solved since many generations of models back. Claiming it's because it has that game in its training data is nonsense.
The goalpost velocity is approaching light speed…
>There's probably a better base RTS on github somewhere for free.
I... I think you are missing the point.
I really wish there was just an easy guide on when to use Sol vs Terra vs Luna, and it just moves further into confusing territory when it comes to naming.
The naming convention is especially difficult to decipher depending on what your native language is. Of course a latin language speaker might be able to easily determine oh yeah each one is slightly bigger than the other but I still think it borderlines too confusing.
That aside all the numbers look amazing, and I'll be happy to probably main this alongside grok-4.5 for a while comparing the two on price and efficiency.
I vastly prefer the direction that OpenAI seems to be going with token efficiency and performance compared to Anthropic who seems to be moving towards a world where you just token-max as much as possible ignoring any and all costs.
Use Luna. It's more performant than 5.5 and it's cheap. Hopefully it's cheap because it's more environmentally friendly than the bigger models. So you're doing a good thing. If it's a smaller model it may even be faster, but I haven't looked into it yet.
In my tests, in almost all cases, using Sol on (low) reasoning is the best option intelligence/price-wise.
Luna is good too, for classification tasks or any pre-processing task that is not critical
My guess is that it's the same for Haiku/Sonnet/Opus: Biggest model for architecture and high level planning and technically challenging problems, medium model for simple implementation tasks, small model is for nothing
it's simple: unless trivial TOIL, always use the highest at ultra max settings.
Non sense, and time consuming.
Okay Richie Rich
Why would you need a guide for that now? We long had to pick different models (and thinking levels) by task and feel.
Previously it was much more obvious which model to reach for depending on your use case because they had the mini and nano naming conventions.
Getting rid of that seems like a step back. Just a personal nit though.
I've seen buzz about this elsewhere as well but to me effort levels seem more like spend limits disguised with another word. I don't think they should even exist.
My guide was to pick the best model on "High" for 99% of tasks.
The naming convention is bizarre and doesn't really mean anything to normies. Trying to pick between "Sol" and "Terra" is like asking the average person if they want the Max or the Ultra chip.
What about Haiku, Sonnet, and Opus?
Also just as confusing. Shouldn’t take away from the point though. Both can be bad names.
The sun is bigger than earth which is bigger than the moon, it's pretty simple really
Which is cheaper to use? The size euphemism is a really roundabout description vs "Nano" and "Pro" for the layperson.
You don’t know what sol means? You don’t understand the difference in sizes between Terra and sol? I’m genuinely asking.
That isn't what "genuinely asking" looks like, you're criticizing using "questions" as cover. It isn't subtle, nor is it constructive.
I agree with them, Sol, Terra, and Luna are confusing names. They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
It feels like it was adding by the marketing department.
Surely it's size based: Sun (Sol) > Earth (Terra) > Moon (Luna)
Similar to Anthropic's size/length based naming: Opus > Sonnet > Haiku
These names seem easy to understand to me, and much clearer than suffixes like -max and -plus.
I'd agree it is similar to Anthropic's naming scheme, which I'd argue shares the same problems as this. It improves marketability/googlability, but decreases actual comprehension.
You don't actually explain why or how these names are "easy to understand" just state that they simply are. That's great; to me, they aren't obvious or intuitive at all. May have well just start randomly pointing at dictionary words.
I collect roman coins, with latin legends, so the sun/earth/moon references jumped out at me, and partly based on the opus/sonnet/haiku precedent I assumed that these names were referring to different model sizes/prices in a way that mapped to the names (Sun > Earth > Moon).
I'll admit though that until recently I never really thought about Anthropic's naming scheme as having meaning (an Opus being longer than a Sonnet, being longer than a Haiku).
>They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
But do they though? When do you use GPT-5.6-Max-Low vs. GPT-5.6-Plus High? Or GPT-5.6-Fast-Xhigh? What's the Pareto optimal choice (outcome and price)? According to the benches it seems to bop around and the even if the benches are accurate the best choice isn't always consistent.
> When do you use GPT-5.6-Max-Low vs. GPT-5.6-Plus High?
You don't, because that isn't something I proposed using for model naming.
I called them GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast. Reasoning levels are distinct from the model design itself, and the UI makes that clear.
Plus, using that same flawed argument this would be called GPT-5.6-Sol-Low or GPT-5.6-Luna-High which also makes no sense/is confusing. So that argument applies (or more accurately doesn't), no matter the model names.
I do know what Sol/Terra/Luna mean, but was also confused for a second on the hierarchy. After doing a bit or research it dawned on me that they are arranged in the order of the sizes of the celestial objects but it somehow wasn't immediately obvious to me from the start.
Anthropic ships models with a helpful one-liner tag that makes the model hierarchy obvious. I think it wouldn't hurt if OpenAI did the same.
Sure—so, is Sol 109.2x better than Terra? Or 1.304x10^6 better?
Did you not read the second sentence? Obviously I know what sol is given my first language being Spanish. I'm just speaking in a general sense that it can be confusing for others.
I already know plenty who had no clue what the difference between Terra and Luna would be.
My first instinct was Sol > Luna > Terra, since Sol is the farthest away, then Luna, and Terra is the closest. Size was not my first instinct. Or should Terra be the best model because its closest to people, then Luna because there have been people on it, then Sol be the worst because no human has been there?
The naming scheme is too "clever."
The frontier graph on all these benchmark are extremely in favor of 5.6 Sol over Fable, more than the best model comparisons in previous iterations.
I'd like to know how cherry-picked this is, and what tests it performed less overwhelmingly in, but I suppose that info is not going to be on this post.
If it pans out to be as good as it says, that's great. On the other hand, if this model is not overwhelmingly impressive over Fable, I will lose what remaining trust I had in these announcements.
They do disclose that they scored much lower than Fable on SWEBench Pro, which is a pretty high-quality benchmark. I think it's partially just about what they choose to emphasize...
It's worth noting that OpenAI recently came out saying, "We don't think SWEBench Pro is worth reporting any more" - https://openai.com/index/separating-signal-from-noise-coding...
lol
People will downvote you because this comment is "not appropriate" for HN, but there were countless conversations on HN about how important these benchmarks are.
I am literally LOLing at HN right now
SWEBench Pro should be ignored until they fix it or disprove the broken task accusations.
There has been a lot of chatter ever since the Mythos scores had been release that SWEbench pro had major contamination and that Mythos had memorized many questions that lacked the context to be solvable on their own. And now with OpenAI saying a large number of the questions are broken, I think it's worth taking that single outlier benchmark with some salt when the overall trend is that 5.6 is very competitive with Mythos at about half the price.
I totally missed that, because in the charts they showcase for coding, the SWEBench score is not present, they only include it at the end of the post in tables. Hmm.
Great catch.
The SWEBench benchmarks are really gamed at this point and should not be trusted period. The solutions are effectively in the training sets and have been for a while.
SWE Bench Pro is completely different benchmark than SWE Bench (e.g. Verified) suite was. It only copied the name.
SWE Bench Pro is also gamed and shouldn't be trusted.
Then we are left with what? FrontierCode maybe? IIRC that one evaluates not only if tests pass but also code quality - e.g. whether the maintainer would accept the pull request as is.
Didn't they also just post about how SWEBench is broken?
> SWEBench Pro, which is a pretty high-quality benchmark
No, doesn't seem like it
https://openai.com/index/separating-signal-from-noise-coding...
The proof is in the pudding and these benchmark stats will only work for so long before people lose interest.
The charts are also extremely difficult to parse. They seem auto-generated. Dataset coloring is atrocious.
Regarding your main point, yes, I agree. My impression (as someone who uses both Codex and Claude Code daily) is that OpenAI does a fair amount of benchmaxxing.
Anyone else noticed the "Extended: Fable 5 is included in your weekly limit through July 12 blablabla" disappeared from claude code? Did they panic-delete the july 12th deadline ?
I still see in the menu to select the model in the GUI (Claude Desktop, claude.ai etc).
Yes, I noticed this too!
"We've extended usage of Claude Fable" message incoming any day now.
They reset all usage half an hour ago. It's back to 0% per week and session. No specifically Fable related.
I let it work on two features while I was using mostly GPT 5.6 and it has already consumed 10% of the weekly Fable limit on Max x20.
GPT 5.6 on the Pro x5 plan is down to... 100%. It looks like they just reset the usage limits again. And I still have two resets on the bench.
Anthropic is going to have to up their game to compete.
Hahaha seeing this play out in real time is absolutely incredible.
Im here for it, good on Anthropomorphic to feel some heat again after all that drug dealer Fable business.
100. I'm so tired of being treated like a drug addict by them. I'm currently sitting on _four_ "reset vouchers" from OpenAI so I get basically all week to play with 5.6 Sol to my hearts content. The amount of positive sentiment that brings to me should really be a concern for Anthropic who is increasingly alienating me away with their shit strategies.
Or an allegedly even more dangerous model that they refuse to release. What a joke Anthropic has become.
Here are 18 pelicans - six each for Luna, Terra and Sol at the six different reasoning effort levels (plus the price to generate each one): https://static.simonwillison.net/static/2026/gpt-5.6-pelican...
Or if you want to see some in 3D, OpenAI featured a pelican riding a tricycle, bicycle, pony and another pelican in their livestream this morning: https://www.youtube.com/live/Wq45rvPGNHs?t=1070s
Time to dump this test. Probably not a coincidence every version has the same rolling green hills, gradient blue sky, sun in the corner, etc.
On the one hand: yes, pelicans on bikes are definitely in the training set at this point.
On the other hand: the test is clearly not saturated, given that you can see a clear difference in output at the various reasoning levels / model versions.
I sort of agree, but within the same model I expect the reasoning effort to be reflected in the quality of output and that's basically how it played out. When you're comparing different models, then it's just who benchmaxxed the best and there's not a lot of value there.
I don't know. If they were training on this, I feel like they would be able to get the shape of a bike frame right; it's a pretty simple polygon, and a lot of the bike frames that are getting generated would be impossible to steer.
I mean, humans can't draw bikes!
https://themagnet.substack.com/p/why-is-it-so-hard-to-draw-a...
Goodhart's law.
I think the 'pelican test' is becoming useless. It's been around long enough that now I'm sure good examples are in the training data, and hell they might even do some hand tuning to make it do a decent job since they know people will ask about it.
But either way, with no real way to visualize the result of the text it starts with - it will always be stabbing in the dark. It can't understand conceptually what any of it should look like and then refine the SVG to improve it gradually. It just throws darts at a wall and hopes it comes out alright.
I think it's still useful in a "hello world" sort of way. It means you actually tried the new model.
Honestly that's the main value I get from it myself - making a pelican means I have to figure out API keys and how to talk to the provider, or how to run it locally for the local models.
Pelicans, maybe, but the point is to measure how good the "internal visualization" abilities are. Throw curveballs, like a unicorn with a duck bill serving coffee at a basketball court. An elephant playing a piano while its trunk swings a baseball bat at a tiny alien spaceship buzzing its head.
Have them use tikz instead of svg, or have it write code that moves the cursor and draws the thing in paint.
Compositionality and visualization are generally much, much better at each new generation / release cycle.
It's fascinating how well models have internalized visualizing things without actually having joint embeddings / broad multimodality.
Cool. I still find these a useful visualization of some the qualities of llms. Even if they did train for [animal] on [vehicle] svg, it's still nice to see at a glance how the different models and reasoning levels perform. Lunar misses part of the frame, except on max reasoning. While most of the others have a mostly correct bike at all reasoning levels.
I once used something like karpathy's auto-scientist to mutate the prompts and rank them with a vison model. Some of the winners where pretty neat. I think they have a lot more style than the gpt-5.6 ones. https://xcancel.com/xundecidability/status/20449185674144196...
people are saying this is benchmark is saturated but all of these have occlusion issues, even sol max.
A skilled human artist wouldn't have both legs in front of the bike, or a single straight line representing both leg's crank arms.
Yeah it makes no sense at all to dismiss the test, when even the very best examples are noticeably below what a skilled teenager could produce.
Dead internet theory? Semi-random parroting by real people? Or something else.
The quality of sol on effort=none makes me think this test is saturated or they are benchmarkmaxxing this exercise.
Is the direction of the pelicans encoded in your prompt? Curious why they are all left to right with the exception of terra xhigh.
History moves left to right.
Cultural bias.
What's strange with this is the prompt "Photorealistic photograph of a pelican riding a bicycle down a coastal boardwalk, wings gripping the handlebars, webbed feet on the pedals, large orange bill, detailed feather texture, golden hour lighting, shallow depth of field, shot on a DSLR with 85mm lens, natural motion blur on the wheels" produced, well, exactly what I asked it for. I wonder if I tell it then to make it SVG ...
https://chatgpt.com/share/6a5009de-fff8-83ea-98ff-0da17d1d04...
I'm waiting for the day that the "generate a Pelican" test comes back with a SVG-art like illustration of a Pelican equipment case, like a model 1620 or similar.
https://www.google.com/search?client=firefox-b-d&q=pelican+1...
At what stages will models start to internally reflect the drawn SVG and automatically fix their own mistakes?
I assume multimodal models can do it already do it today if constantly asked "make it better"
I haven't tried this in a few months, but last time I tried a loop that rendered the pelican and asked for improvements the results were actually quite disappointing. Be interesting to try that again against GPT-5.6 at Claude Fable 5 though.
They said in the AI community, a pelican riding a bicycle is a good test to measure effectiveness of the model, wondering if they were referring to you, or is it really a standard in the AI community ?
Also would be good to have a tool where users can select models and instantly see each model's generated pelicans. That will make it easy to compare the output of different models.
Simon did start the pelicans on bicycles as an SVG, but I think it's more of a fun goofy thing to see how the model performs at. I don't think it has a direct correlation to a model's performance though.
Surely "how to draw a SVG pelican on a bike" has made it into the training data by now ...
If that was the case in a non-trivial way you'd see mode collapse, but you don't, they come out differently.
It's because all the frequent comment that this pelican is in the trainings set now also got into the trainings set and models adapt. /joking (I hope)
Surely this comment is literally on every new model release post.
It's part of the pelican tradition at this point.
I eagerly await the models replying with that: "I'd be happy to create a pelican riding a bicycle, but just a note that this might already be in my training data. Simon."
Someone told me that a model said "oh, the classic" when they asked for this recently. I think that was Opus 4.8.
And yet... only one is any good.
Ok, I'll never use max effort again on OAI models..
Is that... an x-rated, censored pelican?
something is wrong with Terra model series, most pelicans, except Max, looks bad
Seems to match the pareto frontier on Artificial Analysis as well. Terra is nowhere on it.
Thank you Simon! Luna is surprisingly decent across all reasoning levels.
I think all of Luna's are bad. The only decent one is sol @ xhigh. Even sol @ max is weird. Sol @ high and @ medium are ok, and every other single one across every model is bad.
Strong disagree, but to each their own. For sol I really like how only medium uses the wings on the handlebars to ride the bike. For all the other sols the pelican evolved a new set of arms separate from the wings.
somehow Terra really struggles here even compare to Luna.
Apparently plus users do not have access to Sol, so I'm really worried about the ugly Terra Pelican.
max effort sol clearly over-engineered
gpt-5.6-sol Max pelican didn’t skip neck day
gpt-5.6-sol x XHIGH is my favourite
AI really sucks at bicycles...
LLMs really suck*
I like Terra High the best. That pelican is utterly yoked.
this looks like the same shit from 4 years ago. give it up.
The first time I did this was actually less than two years ago - in October 2024 - and it's fun seeing how much better they've got since then: https://simonwillison.net/2024/Oct/25/pelicans-on-a-bicycle/
Absolutely, the change in quality over time is a great yardstick.
Nice to see you did the quality level comparisons and did three passes.
I've been using that technique myself on my image gen reviews[1] and it also works well in presentations and for personal study.
[1] https://generative-ai.review/2025/12/beast-mode-activated-op...
Things I have been struggling with Fable over and GPT 5.5, were just solved handily by SOL in a real "thank you, next problem" kind of way. Overall, something that just works is way less wasteful for your usage than struggling back and forth for hours.
Unfortunately, I'm finding that in long-form agentic use, when I'm trying to use Sol, I keep tripping guardrails – moreso than even Fable, somehow.
I don't know exactly what part of my codebase is triggering it, so I'm going to have to keep poking, but apparently the guardrails are not that gentle despite the phrasing. :(
Sounds like you are working on something naughty :)
>> approximately 700,000 A100e GPU hours of black-box automated red teaming
Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.
Not nothing either, but far less astounding sounding than 700k hrs.
I'm pretty sure Altman has spoken about giving a model 100k+ A100s specifically, this might be them being very literal
Wait, what do you mean? 700k A100e hours are equal to 200 hours of a GB300 NVL72 rack? One GB300 NVL72, 72-GPU rack has equal processing power to 3500 A100e GPUs?
maybe? ai says about *8.3 days* of continuous runtime on a single GB300 NVL72 rack
about a sprint's level of effort.
a very expensive sprint
> based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point)
The A100 doesn't have hardware FP4, and you'd be running a quantized model with some accuracy loss but unless this was natively trained on FP4*
* to add another layer, they own the model and could apply tons of post-training techniques to reduce that accuracy loss and probably already do
Maybe it’s a bug but on iOS individual paid Pro account - I can no longer see which model is being used nor select which model I want.
"GPT‑5.6 delivers a step change in design judgment. With only high-level direction, GPT‑5.6 creates tasteful, ergonomic, and functional interfaces. Its stronger computer-use capabilities let it inspect and refine the rendered result—not just generate the underlying code or content—so it can catch visual and functional issues and apply finishing touches before handing the work back."
This one is really promising, as it may allow to close major gap with Claude in design/UI skills
Computer-use is a big limitation that my 2015 Macbook Pro cannot handle. I find the Codex cli says it looks at the end output artifact but so often it fails to refine it into acceptable form. If it could use my computer screen and visual inputs for review, it might be able to actually design documents/powerpoints/etc. I'm juicing everything I can out of the 11 year old laptop and I'm honestly impressed at what it can still do.
How dare you point out that 2015 is 11 years ago.
+1. I've been only using Sonnet/Opus these days for UI work because GPT 5.5 just can't do any of that. Its just really terrible. Eager to give this one a try.
Agreed, I’m looking forward to trying it out. I think that the rise of visual design skills that are pretty clearly targeted towards Codex users has lit a bit of a fire under their butts.
5.6 Terra (mid tier model) as good as Fable on DeepSWE while cheaper than Opus API pricing. Seems like a homerun.
GPT usually performs better on DeepSWE while Claude does better on FrontierCode. These two coding benchmarks are pretty much the only ones right now that's still worth taking a look at imo.
DeepSWE seems to strongly, strongly prefer ChatGPT models. There were also major flaws in its methodology pointed out recently, that overlap strongly with the flaws OpenAI pointed out in its SWE Verified report.
I use both ChatGPT and Claude for engineering work on a daily basis, touching performance critical code to application backends to frontend work, and I've found that DeepSWE scores don't reflect my reality when I assess high quality output from the models/harnesses.
Not that Opus always beats GPT 5.5., but that 5.5 is ahead of Opus on a general benchmark smells off to me.
I hope it isnt like Opus eating so many tokens and taking so much time
Really wanna see it in DeepSWE benchmark
I flip back and forth between whoever currently has the more powerful frontier model that isn't cost prohibitive - subscriptions only, API pricing a non-starter. Today that's Fable 5 which has been excellent, as soon as it's Sol I'll switch to that. The OAI/Anthropic harness behavior has mostly stabilized for me with consistent AGENTS.md that I sync with CLAUDE.md - I like pi (pi.dev) and have tried to build it up to get performance comparable to the two "first-party" harnesses, I'm just not there yet.
One major sticking criteria for not going with OpenCode / pi for all of my coding is I want access to the tier-1 frontier model of the day without API pricing - e.g. afaik I can't use Fable 5 via pi harness even though I have a subscription, so for this week I'm on Claude Code. It's not the need to Fable 5 for everything, but even if I just want the marginal intelligence benefit to stress test an architecture decision, it's a safety blanket to know there isn't a ~smarter~ model I could have used. And for my use cases, the doggedness and capability of these frontier models has been insanely effective.
My feeling is we're still in the Uber era subsidy period - the moment the subscriptions either try to lock me in longer than a month or stop OAI/Anthropic stop delivering frontier models in the subscriptions, I'm out - switching fully over to pi.dev or another OS harness and routing my token spend via OpenRouter or offloading to Qwen locally. Then I'll have to put an accurate dollar amount on frontier intelligence.
> My feeling is we're still in the Uber era subsidy period
I often wonder whether this doesn't continue indefinitely.
Uber was able to do this because it was just them and Lyft playing second fiddle, with a huge barrier to entry once the network effects had kicked in.
It just seems like the model space has way too many competitors, + OSS/Local options for them to ever be able to jack up their prices. At least once the datacenter bottleneck has been cleared.
I'm working on a multi-harness IDE that supports custom agent workflows and skills that are shared between any harnesses it wraps over. I think it might prove handy for a workflow like yours.
Would it currently support Fable 5 via the restrictions Anthropic is placing on usage... because that's my major blocker
GPT-5.6 is a really good model, and quite cheap. I can finally replace GPT-5.3-Codex for my Tool Calling in n8n.
Here's my benchmark results for GPT-5.6:
https://aibenchy.com/?q=gpt-5.6
(the high reasoning variants are still running, uploading them soon too)
EDIT: The high variants are there too, enjoy the hamsters[0].
[0]: https://aibenchy.com/showcase/?q=gpt-5.6
Interesting that Sol (low) did better than Sol (medium) in your benchmark (and is barely more expensive than Terra). I too have been using 5.3 codex as a cheap-but-good model and are switching to Terra (xhigh).
Yeah, for some reason the (low) versions do really well, like they think directly of the solution instead of going around all the edge-cases and getting lost in one of them.
Also for most, there doesn't seem to be a big difference between (medium) and (high).
Here's all 3 (medium), and GPT-5.5
It GPT-5.6 doesn't seem to be a lot smarter than 5.5, but it is faster, cheaper, more efficient and more consistent:
https://aibenchy.com/compare/openai-gpt-5-6-sol-medium/opena...
Benchmarks look really promising. Suspiciously good, even. I guess we’ll see soon enough.
My question to previewers: how are the guardrails for random joe that wasn’t personally blessed by the ai pope to access the non-nerfed model? Fable is a nightmare in this regard, but I’m not sure whether 5.6 also gets a critical side-eye from the gubmint when you ask it to fix bugs in your code (you filthy hacker, you).
I almost immediately ran into "This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests."
Which is something I've never seen with codex before, and I wasn't doing anything funky. Just writing CUDA kernels and benchmarks for them.
Is it actually usable though? Because the Fable situation is just obnoxious. If OpenAI's Fable equivalent is actually usable, I'll cancel my Anthropic subscription on the spot.
Yes it's usable, however it does make it difficult to just do /goal and walk away if this is happening periodically. I don't do that often, but for well defined things I sometimes do.
Great. I usually babysit the agents so I think it won't be a problem. So long as there's no obnoxious downgrading going on.
That settles it. Anthropic has until the end of this month to get rid of Fable's "safety" bullshit. If they don't, I will not renew my subscription.
I was getting that regularly last week with regular 5.5 medium on the plus plan. I was doing benchmarking for a photo editor in Swift.
Interesting. Never seen it before. Now it's just constant and there's nothing sensitive about this work.
I'm getting the same message doing WebGL shader work.
I haven't tried an OpenAI model for a long time, but with Fable going to API pricing soon this might be enough to get me to try codex.
Seeing how Anthropomorphic just reset usage quotas back to 0 and the other day extended Fable sub inclusion by a few days, I have a feeling they might not drop Fable out of sub after all, because like you I would most definitely take a long good look at codex at that point.
It's not just the API pricing either, there's also the constant uncertainty. They pull the model then put it back up, they say the model is going away then suddenly it's not. And then there's the fact Fable is barely usable because it randomly downgrades to Opus out of nowhere whenever it thinks about exploits.
It's definitely good that Anthropic's feeling the pressure. Anthropic has worn out their welcome with this "safety" nonsense. If OpenAI actually lets me use the LLMs on a subscription without any of this bullshit, I'll definitely switch.
In the introduction video they say 5.6 Sol autonomously post-trained 5.6 Luna. Curious what this means.
/goal tune 5.6 Luna parameters until performance is maximized across all benchmarks
Sounds like they gave it a goal to hit certain benchmarks and just let it have its way with the base Luna model.
This produced a disturbing mental image.
It means OpenAI and Anthropic are now in a RSI race with each other
Wow the video is much better.. the PR spend clearly went up a lot. Mainly just showing "real people" doing "real stuff".
When in doubt, use grandmas knitting. yeah, that will make them trust us.
I use 5.5 a ton. It's immediately apparent that 5.6 is truly a better model. Hope they don't lobotomize it later.
Based on the Intelligence vs. Cost graph, not clear to me why anyone would use Terra? Luna looks quite interesting though, happy to see OpenAI still serving the more budget-oriented side of the market (seems like Anthropic and Google have lost interest there).
https://artificialanalysis.ai/articles/gpt-5-6-has-landed
Cost and intelligence aren't the only axes. Terra has better latency and output speed than Sol for example.
Very interesting: I wonder if the RL approach is diverging between Anthropic and OAI?
I noticed that Fable uses shell tools almost exclusively (even to search and edit files), compared to previous Anthropic models.
Having run some experiments with 5.6, I notice that it uses built-in file systems and provider native tools much more (not shell tools), compared to previous OAI models.
I can't try it since it hasn't appeared in my Codex yet, but this is is necessary from OpenAI in my opinion. Fable is just so much better at understanding broad context. I only use GPT 5.5 for straight forward easy to describe tasks, and it does crush those. But I spend a lot more time steering Codex towards good design on broad concept type tasks, ones that Fable shows sometimes surprising clarity.
I look forward to seeing how it compares once I have access. Not getting tripped by spurious safe guard flags could be an advantage.
I’m interested in knowing how each of GPT 5.6’s variants fare in non-English writing/translation tasks.
GPT 5.5 has a tendency to write English calques and non-idiomatic prose in other languages. Although that can be somewhat tamed with detailed instructions and a corpus of confusing terms, the model’s output often reads like a literal translation rather than native prose. Since I notice these issues most clearly in languages I know well, it makes me reluctant to trust the model’s output in languages in which I’m less proficient.
Ironically, ChatGPT began as a simple text-generation tool, but much of its offerings and benchmarks now focus on coding and agentic workflows, while leaving behind what made it notable in the first place.
5.6 sol ultra just nuked my branch and burned my 5h limit. nice work
Just used terra ultra for exactly one prompt in codex and it ate through my full 5h window in about 10mns (20$ plan). The results look pretty good though. Luckily I have had my chatGPT subscription for a while and have a bunch of resets available (nice compared to anthropic).
Assuming I take the 5x plan it would give me about an hour of active sessions with terra ultra (maybe ultra is not good value regarding tokens?), not even using Sol yet. Does everyone using codex use the 200$ plan?
I normally use the 100$ anthropic plan and barely ever reach the usage limit.
> maybe ultra is not good value regarding tokens?
Well, yes, as explicitly stated on https://openai.com/index/gpt-5-6/: "ultra goes further by coordinating four agents in parallel by default, trading higher token use for stronger results and faster time-to-result on demanding tasks."
thanks, it makes sense, I'll stick to max from now on
I use the $20 plan, but I don't code all day every day.
With Codex, it is my experience that I can churn through a 5h window in no time with newer models -- especially when they're new. So I tend to use fancier models for planning, and the less-fancy models for writing code based on that plan. I switch to the fanciest model if any part of this gets stuck.
If I've got a something big-ish to work on, I pay attention to the reset timers so I can get more of it done in one chunk.
Models seem to slowly get better/relatively less-expensive as they age. (It isn't clear to me if that's because the cost actually goes down, or if the allotment goes up, or if things get more efficient in unseen ways, or what. OpenAI is vague AF about what we get for the $20 that we pay.)
Do you know if you used sol/terra/luna?
Is any of those comparisons about Pro vs non-Pro (Pro is only available in $100+ plans)? I am curious about that but I think Sol, Terra, Luna are different sizes of it without the Pro part, and I want to know how much worse do I have it on the $20 plan compared to if I upgrade.
Huh, a good alternative just as anthropic's 50% weekly subscription subsidy is ending this weekend. Time to see if it's benchmaxxed or actually a strong leap over GPT5.5.
They also seem to really not care about alignment, or care about it in the wrong way. It's entirely missing in the blogpost and there are some concerning bits in the model card, seemingly treating CoT controllability as something to be "investigated" rather than the warning sign it's supposed to be.
There's also this:
> GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated -- https://www.lesswrong.com/posts/JFjNmPTbH8kL6xtp6/gpt-5-6-th...
it seems terra is pretty much useless, you either want luna max for everyday coding (cheaper and same perf as 5.5 high), or sol xhigh/max for demanding tasks
8% on ARC-AGI-3, they actually got some traction going...
note that ARC-AGI-3 has its rules changed.
before today all the contestants were capped at $10k
So with this release do they kill the 5.5-Pro model with super long thinking and reasoning? 5.6-Sol-Ultra is not the equivalent, right?
On top of GPT 5.6 Sol they added a Tamagotchi / Clippy mascotte https://x.com/giorgio_zampa/status/2075319657997750495?s=20
What in the world is that? Why.
No, this Pets feature has been around for a while now.
I am seeing some bugginess in testing:
Official OAI .NET library. Even when I override the currently experimental [?] flag to 'none', it will still occasionally throw this error (about 5% of the time).I hope we aren't trying to push customers off the chat completion endpoint... Responses endpoint looks great on paper, but the business wants more visibility and control over the reasoning process than this product currently offers.
Edit: This is broken in my VS copilot setup too.
On the tiny voids demo: does your Firefox js thread lock up as well, when you try to interact with it?
https://openai.com/index/gpt-5-6/#a-leap-forward-in-design
Yep, happens to me on Chrome as well
For context, I have access to MS Copilot through my workplace. To see what it looks like, I have tried to login through https://copilot.microsoft.com/ , where I was informed that my account, although recognised, is not yet supported. However, I can get more or less the same chat window, with access to all the data, through https://m365.cloud.microsoft/ A redirect could have been useful.
Dirac (https://github.com/dirac-run/dirac, https://dirac.run/) now supports gpt-5.6. This thing does now seem to be on the chatGPT/codex accounts yet.
UPDATE: it is now available in chatGPT account also, they rolled it out
Will be there soon according to the last commits in the codex repo: https://github.com/openai/codex/pull/31684/changes
Also, confirmed it works for me by using --model gpt-5.6-sol
Does it support subagents?
yup it does
I used to pride myself on not being the "fonts too pointy, scroll too buttery" crowd! But AI has brought me full circle and now nothing removes my interest in reading even a single word on a page faster than purple gradient greeble-afflicted tailwind-slop models put out without stronger prompting/references
That being said, maybe 5.6 can fix that!
Thanks, I needed to hear that lol. Yes, the site was an afterthought, core work took/takes most my focus. I will look into un-slopping the site soon.
One of my best use cases for the short duration I have fable is to use it to create the plan and acceptance test files then use GPT 5.5 Pro to do an adversarial review on the plan then feed that feedback into fable to fix the plan.
Zero information on the knowledge cutoff. The model itself responds it's June 2024 which is weird given that GPT-5.5 has knowledge cutoff at August 2025.
So glad Fable limits just got reset. Thanks OpenAI.
Oh hey, thanks for the hint!
> Instead of requiring developers to script every step or passing every tool response back through the model, Programmatic Tool Calling in the Responses API can filter large amounts of intermediate data, retain only what matters, and adapt its workflow along the way.
this seems very interesting
There is an issue on the page that causes the benchmark tables to get cut off. If you highlight and drag right you can see a few more models like Gemini and Claude Opus. It's also interesting that they introduced explicit caching, which is something that only Anthropic had for a long time.
Looks like I have access to gpt-5.6-terra and luna. How does one decide between gpt-5.5 and gpt-5.6-terra? Pricing is similar, but it's hard to tell if it's better..
this is exactly my question. I would expect that luna is analogous to mini before, but is terra equivalent/better than 5.5 and Sol is a step above? or is terra nerfed and 5.5 is analogous to sol?
Maybe Terra = mini and Luna = nano?
I wish model launches were like proper product releases
it's impossible to _try_ it out on release!
it's not on their codex subscription, or the web/mobile chatgpt interfaces, or aws bedrock, etc. I just cant find a working endpoint with the latest model after they announce
For me, minutes ago, as a Plus subscriber:
I started up Codex CLI fresh. That version of Codex was 1.42.5. 5.6 wasn't in the models list.
After I updated Codex to a newer version (0.144.0), 5.6-terra and -luna appeared in the models list (but not 5.6-sol).
(It's impossible for me to know whether updating was causative or just correlative, but that's the timeline I experienced.)
The announcement says they're rolling it out over the next 24 hours or so. I think it's reasonable to do a slow-roll-out release for one of the most used products on the internet.
GPT-5.6 Terra just showed up in Codex for me.
Looks like a great set of models, but there are about 20 different thinking/model levels here in this family and they are very complex to pick the right one for the task
E.g. for GeneBench Pro, it looks like you would always use GPT-5.6 Sol over Terra/Luna, its pareto optimal.
For Agents Last Exam, you would maybe want Luna, then Terra, then Luna, then Sol as you increasingly budget for tasks.
I feel that there may need to be a new auto mode in many of these cases. It selects the best model and thinking given a particular problem.
Feels like it's going to have to go that way eventually, because here we have about 20 different model and thinking levels you could use, and they're not obvious which ones are right for the given use case.
I think 5.6 Sol is only as good as 5.5 or Opus 4.8 in terms of getting its given work done. It just has an uncanny ability to pickup more work that it can tackle next that the older models lack, or have not been trained to do before. Where folks are seeing a difference between working with Fable or 5.6 I think also boils down to this phase shift.
The claims are pretty bold. I think 5.6 may exceed Fable.
Not specific to OpenAI / Codex, but I'm curious what people are doing to protect themselves from any destructive actions by their coding agents? Just install and pray? Explicity approve all actions? Reconfigure for safety? Run in a sandbox (Docker) ?
Typically I just want to isolate the agent disallowing it from accessing other parts of the filesystem. Using a different user might be enough, but I typically use [bubblewrap](https://github.com/containers/bubblewrap).
I live in fear lol.
More seriously, I was blindly trusting the auto-classifier from claude code (same as the middle option when you do `/permissions` in codex), and it actually allowed the agent to do pretty hardcore `rm` and `git push --force-with-lease` commands, which I would have expected to have to approve manually. Luckily no major issue from those yet.
The best option imo is the integrated cloud environments from claude code (not sure yet if there's a codex equivalent). It spawns a VM in the cloud where the agent runs, and you can open a PR from the app when it's done. Very smooth experience
Interesting - I'd never heard of this Claude Code VM option.
Does it auto install all the dev/test tools it needs, maybe including things like web server & browser? Does your code live in the VM, or in some external repository? Is the lifetime of the VM the same as the agent, or does it persist until you remove it?
Where can I find documentation on this?
I use it for basic web stuff so haven't pushed it to the limit, but they have tons of stuff you can configure (up to self-hosting your environments): https://platform.claude.com/docs/en/managed-agents/environme...
Thanks!
I use the auto-reviewer for actions outside the builtin sandbox.
So far this has been rock solid, and tens of millions of developers use this setup without issue.
It is not going to wipe our hard disks. At least I hope so. Fable and GPT 5.6 have been ever more proactive, and GPT 5.6 is automating the AppStore on my machine to download an Xcode update while I am typing this.
Is this auto-reviewer part of Codex? Is the review done by the agent or the model?
Yes. In Codex it is called 'Approve for me', in Claude it is 'Auto mode'.
I believe in both cases it is prompting a model with a fresh context that is tasked with reviewing the reason for the action.
With Claude, I have seen that if the reviewer does reject the proposed action, it responds with a long text about how the Agent should not try to work around this rejection, and instead prompt the user for an explicit approval of the proposed action.
I still just explicitly approve all actions and review all code (unless it's a personal/throwaway project no one else will ever touch/use/see). I know a lot of people that run in a sandbox though. That said, I'm sure there are lots of people that just yolo it and hope for the best.
What destructive actions are you afraid of in particular? Honestly the models are pretty smart, I let the agents go --yolo and nothing bad has ever happened (yet) that couldn't be solved with git.
I'm not concerned about the code it's working on, but rather anything else - modifying files outside of the project dir (e.g. incorrect tool call), modifying system configuration, doing something bad on the internet, etc.
It is a valid concern but I've been running yolo mode since the inception, using Claude and now Codex. I'm not bragging or anything, since I'm feeding my own curiosity too, trying to answer what's the worst that could've happened? So far, nothing catastrophic that I recall.
Don't let it outside the sandbox. Don't let it have access to anything but dev environments. Continue using git.
Never had any issues.
Which agent/sandbox are you referring to?
good alternative, while gemini still no news
I guess Plus accounts don't get access to Sol? Or is it because I am in Europe?
Oh man, I love capitalism spoiling us here. I was just enjoying my extra Fable credits, now I'll switch to using 5.6 this weekend. I was planning to ration my Anthropic credits, I guess now I do not have to. And I was half wondering if exactly this would happen: right when Fable usage credits were starting to kick in for people, OAI swoops in and takes the puck. As much the AI craze is crazy, this play by play part is pretty fun.
top it off with anthropic stressing about the release and resetting usage to 0 for the week just now.
Anthropic just reset all limits, including Fable. Capitalism is spoiling us.
Make hay while the sun is out.
If hoarding is spoiling. You know what would be better than using fable and gpt 5.6, being able to run that level of model on your own hardware.
Will this run on Cerebas? I'm really looking forward to that.
Sam Altman confirmed during the initial limited release that Sol will run on Cerebras at 750 tok/sec.
"I canna' give her any more, Captain!" - Montgomery "Scotty" Scott, Chief Engineer
This is the part I'm most excited about with the new release, though I'm concerned plebs like me may never get a chance to play with it
we probably need to use gpt sol max to decide which gpt flavor and effort we need to use per task.
We have an official pelican on a bicycle from the OpenAI livestream:
https://imgshare.cc/mz9xwut3
holy moly it's in THREE dimensions!
AGI solved
So it's failing epically because it generated a tricycle instead of a bicycle?
The livestream presenter goofed and said the test is typically a tricycle, so I wonder if that's just a coverup.
The site they showed had the pelican riding a number of things, including a bicycle and another pelican (no, not in that way).
I never have have the issues most people talk about ... I feel like most were never Devs before ai and don't know what they actually need done when prompting. that on top of not utilizing good tools such as a codebase indexer, lsp and a project scaffold.
Wow, the "Agents' Last Exam" graph looks unreal!
I mean the y axis is deceptive to make it seem like greater gains since it starts at 30%, when in reality the differences aren't great.
Even worse, it's not a fair comparison: they purposefully just used "adaptive" instead of "max" for Fable.
What about the graph looked so unreal to you?
That’s because it’s bullshit
I find that 5.5 gives me far fewer refusals than Anthropic models for security and reverse engineering work. I hope the same is true for 5.6.
Yeah, I pretty much had to switch to using GPT rather than Opus completely for all my security benchmarking and harness development. I was annoyed enough to blog about it: https://swelljoe.com/post/why-i-had-to-switch-to-gpt/
GPT-5.6 Sol, Terra, and Luna. at this rate GPT-6 will be named after a parking lot and GPT-7 after whatever Elon names his next kid.
From what I see on ArtificialAnalysis.ai, OpenAI cranked up the reasoning effort to surpass Fable, and now the cost/task for 5.6 max is closer to the Claude Opus 4.7 and 4.8 cost/task.
If it's not dangerous enough to be classified as WMD by USG, who's interested.
does anyone on chatgpt business plan (not enterprise) not have access to the Sol models in codex? i have 5.6 for terra and luna but not sol
The cost & output token charts are useful but I wish I could view them more like a 3D surface. Like the CS:APP memory mountain charts.
I wonder how long model size and effort will be a few discrete points instead of continuous.
GPT‑5.6 system card https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
I am on Plus subscription and see Terra and Luna in Codex, but no sign of Sol. Will it be available only on Pro plans?
I an on Pro and it still returns "The 'gpt-5.6' model is not supported when using Codex with a ChatGPT account"
UPD from announcement: "The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
My Codex app got upgraded to the new unified ChatGPT app. I don't see Sol available though. Only Terra and Luna. I'm on the Pro plan. Anyone else see it?
Same here (Business Plan).
Same, no Sol (i'm on plus)
This marketing video on the page is nice!! can't wait for the hardware to get cheaper to live the AI life i wanna live.
i wish they had renamed chatgpt to codex instead of the other way around ...
Is it available in EU? I only see 5.5 still :-(
Arguably not in the EU, but I'm seeing Sol,Terra and Luna on my account here in Norway
prompts -> loops -> slingshots?
Its an extremely capable model. I think the way we need to approach works shifts again. We need to get our harnesses/workflows to let it gather some momentum on the first couple rounds but then we also need to structure it so that it can slingshot and accomplish the long range goal.
I wish they had kept their previous sensible naming convention instead of this celestial Sol, Terra, and Luna mumbo-jumbo
I assume they're jealous of the Fable/Mythos hype. People talk about Fable like it's a whole new thing, rather than another incremental improvement over the existing best models (which has happened several times and continues to happen).
Overloaded in Codex, no indication if it is already in ChatGPT and I can't use it in the API even though it says it should be available. Typical horrible OpenAI launch. Glad that Anthropic just reset the rate limits so I will go back to Fable again.
I wonder what increment of progress will be achieved by the next billion dollars
They have a fantastic media team.
So is 5.5-Daybreak still relevant for cyber security give. 5.6 capabilities?
>Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
Sounds great.
Also latency looks very good.
Sounds like a perfect fit for a minimal or bespoke harness?
GPT 5.6 Sol is a token hog. After implementing the task, it started some "reviews" I didn't ask for - they consumed 19.5M and 11.9M tokens, while the task itself was below 5M tokens.
> GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints (opens in a new window) and a 30-minute minimum cache life.
Great to read they are moving away from the 5 minute cache defaults. Hopefully other providers follow soon!
They highlight the cache write price now much more in the guide. Did it increase vs. prior generations?
There was no cache write before!
https://openrouter.ai/openai/gpt-5.5?endpoint=58e5b336-423e-...
vs
https://openrouter.ai/openai/gpt-5.6-sol?endpoint=a54c5de0-8...
Almost immediately ran into some the kind of gatekeeping I've heard Claude Code users complaining about with Fable. Not sure why, I just had it working on writing benchmarks for some CUDA kernels. Nothing security related:
"This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests."
At least it gave me the option of waiting instead of just unceremoniously downgrading me. Appears to be making progress but... weird?
I think the most interesting part of this is that OpenAI is going way easier on the classifiers than Anthropic. They explicitly state that many defensive cybersecurity uses are supported and implicitly criticize Anthropic's stance on Fable's uses by saying that overblocking cyber requests is itself a major security risk as more AI models continue to advance in intelligence. I have so many questions as to what is going on on a game theoretic level in the AI space in the past two months, it seems like multiple actors have realized their incentives are really quite different than they originally thought.
It's good to see labs taking into account the cost/task.
Grok 4.5 is interesting because it's smart enough at great price. It seems gpt 5.6 is right there with great efficiency and great pricing.
Working with Fable has been a great experience, but at the end of the day, if you can get only 10% of your work done because it just burns through tokens, that's not that interesting.
I've been mostly using Opus and Fable high for planning and codex 5.5 medium for implementations. Claude is also the only model i can use for design tasks. If gpt 5.6 can finally deliver on the design side, it might be time to ditch the Claude sub and go full Gpt.
Using the Claude "superpowers" skill will downgrade models automatically, using Sonnet and Haiku for trivial things.
For writing GPT which i was subscribed to Fall 2024 to March 2026 (laid off) is superior to Gemini. Been using Gemini since March mostly and they offered a $10 a month plan so i took it. Though today realizing GPT is superior to help me write I am back to being a paying customer. Im in full swing mode to get back into the job market (get the heck away from UI/UX which is now a stupid career in terms of number of jobs out there and in the future there will continue to be less) pivoting into product management (can vibe code anything now) and or customer relations. Hopefully GPT helps me with this pivot and Im again gainfully employed!
Bro these colors on chars are unbelievlable, I can not understand which is opus, which is fable, which is GPT...
GPT Terra is 50% cheaper than 5.5 while being more performant. So it’s like a straight up 50% reduction in cost!
That leads me to a question. Why wouldn’t they just default to terra in ChatGPT in the last few months? If they didn’t then they burnt money for no reason by giving a shittier model at a higher price
"while being more performant"
..on some specific set of benchmarks ;)
where is it? Still not accessible...
"GPT‑5.6 is available starting today across ChatGPT, Codex, and the OpenAI API. The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
cursor benchmarks with GPT 5.6 in picture, a good reason to stop using opus.
https://cursor.com/evals
The good news you don't have to send your dollars to China to fund ai dictatorship, in russia, north korea, african countries and south america.
so the answer is use grok ?
I'd say answer , the opus is no longer undisputed. grok + gpt models are very competitive + glm if you are ok to wait 3-4 times longer, unless you have some unique access to GPU
If Fable is removed from my Anthropic sub, I'll have to change to OpenAI.
"On Agents’ Last Exam (opens in a new window), an evaluation of long-running professional workflows across 55 fields, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable: GPT‑5.6 Terra and GPT‑5.6 Luna outperform Fable 5 at around one-sixteenth the cost. "
Some pretty big claims and results! Excited to see how it feels during usage.
I use Fable and 5.5 extensively and I still find both have a place in my toolkit, i.e. Fable IS good but it isn't perfect, and it's still better to play them off against each other. I have Fable and 5.5 write plans and have them adversarially review each other's plans.
Having this amount of competition in the coding model space is good for all of us.
I think this is the phase shift 5.6 (Sol set to Ultra) is bringing to the table. Until now we have become accustomed to asking models to continue and their natural inclination is always to stop. Now OpenAI have flipped it around and for the first time are asking us to steer or stop the model instead, and its own inclination is to keep going. We now have to decide when we need to steer or want to catch up on our understanding of the work done but it will keep going.
Benchmaxxed
Here's me using a Gemini chat log scraper (from Gdrive) then dumping my prompt+Gemini response into local AI
Never go over the free limits in Gemini Pro.
Gemini is great at research and architecture, and my 30 years experience in programming everything; for fun or work; means together there is little to no code slop.
Add to project repo some git submodules of reference source code; boom, bobs your uncle
Zero reason to sign up for OAI or Claude. With employers realizing the costs are more than employees, local models getting more powerful, and models in chips just a few years out, neither of the one note LLM companies without diversified services and R&D portfolios gonna last
i'm not happy with how openai is trying to pit 5.6 sol as a cheaper equivalent to fable here
for one thing, they said that on AA, sol is "within one point of fable" at 58.9 vs 59.9 but don't clarify that the latter is with safeguards where ~8% of the tasks got routed to opus
i'm not rooting for either and genuinely think that the token efficiency and cheaper price are important but this sort of thing just feels disingenuous :-/
This is especially interesting because IIRC the AA benchmark is calibrated so that 1 point and greater difference is statistically significant.
Like the last time, again they failed to note whether there is an Instant model or when it might become available.
Just a day before my $100 subscription expires, perfect
Do they expect us this model is 15ppt more accurate at half the price of fable? What’s going on?
Not available - checked and it's not there.
As usual, even though GPT-5.6 is releasing today, the rollout in ChatGPT and Codex will be gradual over many hours so that we can make sure service remains stable for everyone (same as our previous launches). We usually start with Pro/Enterprise accounts and then work our way down to Plus. We know it's slightly annoying to have to wait a random amount of time, but we do it this way to keep service maximally stable.
The timescale is typically hours not minutes, so if you don't see it now, I'd try again later today.
We mention it will be a gradual rollout over the next 24 hours in the Availability section at the bottom of the blog but I admit it's pretty buried.
(I work at OpenAI.)
Is this bug fixed with 5.6? If not, it probably doesn’t matter which version Codex users are getting because the overall result is dramatically worse than stated by Open AI advertising: https://github.com/openai/codex/issues/30364
Understood thanks; will 5.6 fix this issue that makes Pro unusable?
https://github.com/openai/codex/issues/30364
"GPT-5.5 Codex reasoning-token clustering at 516/1034/1552 may be leading to degraded performance on complex tasks"
on Plus I see Terra and Luna, but not Sol
It's available in Cursor now.
They have really been stringing us along for the past few weeks.
If OpenAI can add all the features from CC into Codex i’ll gladly switch.
Which features?
You posted the same comment twice.
Thought Fable was great
good alternative to anthropic
I'm disappointed these models continue to be closed source and so expensive.
Open weight models being 10x or more cheaper is just so much more of an unlock than incremental gains for me.
because they're stealing from the frontier models. they're gaming the benchmarks. look how bad glm 5.2 is on cursors evals. gmhit garbage , but it gets glazed as God tier.
They're stealing, eh?
The meat of the report for SWEs:
SWE-Bench Pro Sol: 64.6% Fable: 80% Opus: 69.2% (!!!!)
So, it still trails Opus, significantly, and is not a next-gen coding model like Mythos/Fable 5.
Disappointing to say the least, but somewhat expected.
SWE-Bench pro is pretty much useless now even though many ppl still look at it. OpenAI published a report yesterday saying so as well. Only look at DeepSWE and FrontierCode right now for coding imo.
Amazing, a company that does poorly in a benchmark says that benchmark is useless...
SWE-bench series just aren't that great by today's standard, even Anthropic previously stated Claude had memorized solutions for the non Pro version of the benchmark, I suspect the recent increase in the score for the Pro version probably also had similar behaviors.
But anyway, I think it's pretty useless to look at SWE Bench's now when other way better benchmarks exist.
> Anthropic previously stated Claude had memorized solutions for the non Pro version of the benchmark
yeah that was the point of introducing the Pro version
You've overstated the conclusion. The SWE-bench series has had issues since its inception.
OpenAI no longer recommends SWE-Bench-Pro as a benchmark: https://openai.com/index/separating-signal-from-noise-coding...
That's smart, they only recommend benchmarks that make them look better then their competitors.
Makes sense why they released an entire study yesterday discrediting SWE-bench Pro.
And they'd be right, it's an almost saturated benchmark where even some subpar open source models score very well on. And most models are clustered within a small range so it really doesn't tell you much.
if OpenAI adds all the features from CC into Codex, i’ll gladly switch.
What features are you missing? That you can't add through skills?
> On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol with max reasoning sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less.
> That advantage extends across the family: Terra performs just above Fable 5, while Luna outperforms Opus 4.8; each does so in roughly one-third of the time, with about half as many output tokens, and at approximately one-quarter the estimated cost.
Wow. I don't believe it. Every indication and twitter post told me that Fable is much more intelligent than Sol and here we are told that even Terra outperforms Fable?
Not only that, Sol doesn't even come with run time classifiers. So it is even more suspicious.
What's even stranger is that OpenAI is directly referencing a competitor in this direct way.
CTRL-F: Fable
15 hits
Holy shit. They must be feeling very threatened by Fable if they're spending this much energy talking about it in the release notes for their own model.
In the past they received a lot of hate for not comparing to the competition.
yikes - looks like you need to go back to stats school
gemini - 13 hits
opus - 18 hits
So they are more threatened by opus than fable, or are they almost as threatened by gemini as they are by fable?
The second paragraph has four mentions of Fable. I think that makes my case pretty clearly.
Anthropic fumble of Fable's release will go down in the history books, makes sense for OpenAI to run with it.
Apparently it significant outperforms fable on both an intelligence and cost index.
I don’t believe it at all and I don’t think anyone else does either.
I believe that it outperformed it on benchmarks.
Downvoted comment but I did find this comparison aggressive and tacky.
At this point, they are just changing the decimals to stay relevant and in the news.
Anthropic should be grateful OpenAI did not borrow "Epic" and "Legend".
I expect OpenAI names to be "fabulous", "glorious", "empowered", "delicious" etc.
Most importantly, the cost:
> GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output.
Just as expensive as Fable 5. But of course, another slot machine upgrade but the costs will keep going up and the open weight models from china will continue to race everyone else to $0.
Looking forward to the next version of GLM, Qwen, Deepseek and Minimax.
That's wrong. GPT 5.6 Sol looks to have the same price as GPT 5.5, apart from a new pricing fee for cache writes. Fable 5 is $10 input / $50 output.
Also watching deepseek closely. Seems like US frontier labs only know how to throw money at things as opposed to actually make smart improvements to the algorithms.
To be fair, DeepSeek doubled prices during the peak Chinese workday. (Which admittedly doesn't affect me much.)
The marketing team must've done research that said "people are starting to think that you guys are evil-water-stealing-lay-off-loving-bubble-bursting scumbags" and decided to really lean into the small family business and happy font vibes!
The way they talk about cyber security fixes makes clear that they are in bed with the government in order to get ahead of Anthropic.
All of them closely collaborate with the government. LLMs are a national security priority and are vetted. Claude AI was used by Palantir's Maven to target the Minab school that led to a triple tap strike killing over 150 schoolchildren.
The Minab disaster has every sign of being a pure humint fail the defense department decided to cover up with politically expedient AI blaming.
Claude makes suggestions for targets and humans review and approve them. We always knew there was a human in the process but if there's one massive takeaway from years of AI ethics research, it's that there is a very clear and well-documented human bias towards automated answers when there's any ambiguity.
Including a human in the loop does not excuse the fact that AI was trusted in a process that decides who lives and dies.
The DIA's Maven database was out of date: https://www.theguardian.com/news/2026/mar/26/ai-got-the-blam...
"We triple tapped a girls elementary school because our data wasn't up to date."
Weirdly, normally new ChatGPT releases are head and shoulders above anything else, but according to OpenAI's own evaluation, Anthropic's Mythos outperforms ChatGPT in quite a few benchmarks: https://openai.com/index/gpt-5-6/.
ChatGPT 6 must be deep in the pipeline and will be released within the next few months. Maybe that's why this release is versioned 5.6, not 6.0.