I used Claude Code a lot until this weekend, when I gave Codex CLI a try, and I have to say, wow. The gpt-5-codex model is amazing. Sonnet 4.5 routinely gets stuff wrong, even Opus 4.1 isn't too amazing, but GPT 5 Codex just one-shots everything.
I've been using Sonnet whenever I run into the Codex limit, and the difference is stark. Twice yesterday I had to get Codex to fix something Sonnet just got entirely wrong.
I registered a domain a year ago (pine.town) and it came up for renewal, so I figured that, instead of deleting it, I'd build something on it, and came up with the idea of an infinite collaborative pixel canvas with a "cozy town" vibe. I have ZERO experience with frontend, yet Codex just built me the entire damn thing over two days of coding:
It's the first model I can work with and be reasonably assured that the code won't go off the rails. I keep adding and adding code, and it hasn't become a mess of spaghetti yet. That having been said, I did catch Codex writing some backend code that could have been a few lines simpler, so I'm sure it's not as good as me at the stuff I know.
Then again, I wouldn't even have started this without Codex, so here we are.
It's interesting how different the subjective experiences of similarly-capable coding models is. My experience with Codex is that it tends to run off and do things without asking enough questions or keeping me in sync, whereas Claude seems to be more careful to clarify and keep me apprised of what it's doing.
I wonder how much of it comes down to how models "train us" to work in ways they are most effective.
I think a lot of it, Claude is definitely careful and Codex runs off too eagerly before discussing much (and the lack of a plan mode doesn't help), but I think we just learn how to use them. These days, anything I don't like goes into the AGENTS.md, where I tweak the instructions until the model understands well.
Weird, I tried the CoPilot and Codex CLIs and my experience was not good. I set it up with the same MCP tools I use elsewhere and the results were subpar compared to using agents in IDEs. I don't think it's a context issue either.
I’m not sure about op but codex high is slow but really solid. Med is definitely more hit or miss. I think the “meta” is claude does straightforward “shoe tying” code better and codex does more complicated thinking stuff better. Especially high.
Tool use codex is trash compared to sonnet. So still not a one stop shop.
After all these years, maybe even decades, of seeing your blog posts and projects on here, surely you must have had more experience with frontend than ZERO since you first appeared here? :)
I wrote a long comment but refreshed by accident before I could post it, so here we go again:
I'll write a post when I finish Pine Town, but I don't know what I could say about Codex in it. I think a big issue is that I don't know what others don't know, as the way I use LLMs (obviously) feels natural to me. Here are some tips that you may or may not already know:
* Reset the context as often as you can. LLMs like short contexts, so when you reach a point where the information has converged into something (e.g. the LLM has done a lot of work and you want it to change one of the details), reset the context, summarize what you want, and continue.
* Give the LLM small tasks that are logically coherent. Don't give it large, sprawling, open-ended tasks, but also don't give it chunks so tiny that it doesn't know what they're for.
* Explain the problem in detail, and don't dictate a solution. The LLM, like a person, needs to know why it's doing what it's doing, and maybe it can recommend better solutions.
* Ask it to challenge you. If you try to shoehorn the LLM too much, it might go off the rails trying to satisfy an impossible request. I've had a few times where it did crazy things because I didn't realize the thing I was asking for wasn't actually possible with the way the project was set up.
That's what I can think of off the top of my head, but maybe I'll write a general "how to work with LLMs" post. I don't think there's anything specifically different about Codex, and there must be a million such posts already, so I don't know if anyone will find value in the above... For me, it Just Worked™, but maybe that's just because I stumbled upon some specific technique that most people don't use.
my issue with codex is it will decide to take forever and do to much for one line changes I should've done myself, and sometimes would make more changes than desired. Claude Code is much more expedient and keeps its scope narrow and rarely goes outside the bounds of my request.
It's really easy to steer both Claude Code and Codex against that though, plop "Don't do any other changes than the ones requested" in the system prompt/AGENTS.md and they mostly do good with that.
I've tried the same with Gemini CLI and Gemini seems to mostly ignore the overall guidelines you setup for it, not sure why it's so much worse at that.
I agree with this, I've hit it too, plus I hit Codex limits in a day whereas I haven't hit a Claude limit yet, but all of this is more than compensated for by the simple fact that the code that Codex writes will almost always just work.
Codex attempts to one shot for me but there’s many rounds of refinement. I haven’t used it in the last couple of weeks because it’s disappointing. Over hyped. Gone back to Amp and a little bit of Cursor with Sonnet 4.5
This is my entire problem with Codex - it will spend ten minutes trying to one shot a problem and usually go off the rails at some point, whereas Claude seems much better at incrementally finding the right solution with me.
I've heard this from many people, but I really haven't had this experience. Sonnet will write code that doesn't work, but Codex will give me working code basically every time. It does take longer, and it does think a lot, but I've never seen it go off the rails.
I do look at the backend code it writes, and it seems moderately sane. Sometimes it overcomplicates things, which makes me think that there are a few dragons in the frontend (I haven't looked), but by and large it's been ok.
If I'm doing a large task, I use GPT 5 Pro to write a spec first (with advice for Codex, broken down task list, snippets etc). I may also supply entire files/repos as context for 5 Pro to produce this.
If I skip 5 Pro but still have a large task, I have Codex write a spec file to use as a task list and to review for completeness as it works.
This is how you can use Codex without a plan mode.
Well, when you use GPT 5 Pro Mode it can't make any code changes, so not really a problem :)
I have similar workflow as parent, GPT 5 Pro for aiding with specifications and deep troubleshooting, rely on Codex to ground it in my actual code and project, and to execute the changes.
Codex won't read as much of your code as 5 Pro will (if you give it the context), and Codex will skip over reading in context that you give it (5 Pro can decide what's relevant after reading it all).
Yes Codex is still very early. We use it because it's the best model. The client experience will only get better from here. I noticed they onboarded a bunch of devs to the Codex project in GitHub around the time of 5's release.
> and Codex will skip over reading in context that you give it
That hasn't been my experience at all, neither first with the Codex UI since it was available to Pro users, nor since the CLI was available and I first started using that. GPT 5 Pro will (can, to be precise) only read what you give it, Codex goes out searching for what it needs, almost always.
That’s what I’m saying. Codex will search but then won’t read full files and is stingy with ingesting context. 5 Pro will take in a lot more context (quality up to about 60k input tokens) but you must give it. So sometimes you can even use Codex first to find what full files you should give to 5 Pro to create the spec/task list.
What my quote meant is that once you have the context Codex needs to do its work, if you give it to it, it’ll start the work right away without going and reading all those files again, which can help minimize context use within a Codex session (by having 5 Pro or just another Codex read in a lot of context to identify what is relevant for Codex instead of having Codex waste precious context headroom on discovery in a session that is dedicated to doing the work).
Yeah I was using Claude pretty continuously for 3, 4 months and then decided to give Codex a whirl and it was impressive. I'd consider it to be a lot more cautious and careful and less lazy?
It is however slow, and more expensive. You can either pay the $20 and get maybe 2 days of work out of it, or $200 for "Pro." But there's nothing inbetween like the $100 USD Claude Code tier.
Yeah, I'm really missing the $100 tier. The $20 gets me a day of coding a week with it, which is way too little, and $200/mo is too much for hobby projects.
I've personally been running the Claude Code tool but pointed at DeepSeek's API platform. Cheaper than both Anthropic and OpenAI, and about as good as Sonnet 4 was, I'm finding.
Context window is too small though, and it sometimes has problems with compacting. But I was having that with Sonnet 4.5 as well.
I really loved using Claude. I like working with Claude more than GPT or Gemini. Claude is to LLMs what Firefox is to browsers. I just like Firefox more than Chrome. It's very clearly behind GPT Codex at this point though. So far I've found Gemini for front-end design work to be better than the others, and I pair it with GPT for everything else. Hopefully Gemini 3 is a solid improvement, I like having at least two LLMs at high quality to run against each other.
There are tools like claude-code-router. I've gone through the pain of getting gpt-5, gemini-2.5-pro, and other models wired together. The system prompt differences are too much though I think, claude still feels the best in claude code.
I'm at the point where I have so much built up around claude code workflows that claude feels very good. But when I don't use them, I find that I immensely prefer gpt-5 (and for harder, design influencing questions, grok-4 heavy which is not available behind an API)
Yeah, I think the system prompts are so optimised for the specific model that others won't work as well, so it kind of defeats the purpose of being able to plug your own model in. I wish I could, but I know I won't get as good performance as with the model's native cli.
I like the workflow with Codex more. Though I like working with Claude more. So I wish Anthropic would copy the Codex workflow.
I like that Codex commits using your identity as if it was your changes. And I like that you can interact with it directly from the PR as if it was a team member.
I'd say this method of coding agent interaction is likely a strong contender for integrating coding agents into teams. You start with a really well defined ticket and a good source of relevant documentation for the project then set the agent loose by assigning it a ticket. It does it's thing, maybe asks questions on a group chat or in the ticket, and eventually produces a PR for the ticket. It's the 'interface' behind how a developer interacts with a project already. There's a lot of hand-waving in there and it's not a today or tomorrow thing, but it seems like it's coming fairly soon.
I use it (and Codex web) specifically when I'm not at my desk (or I am but in the middle of something else) and I want to do something fairly speculative. Kinda either exploratory or investigative. I may or may not use the results but it doesn't get in the way of anything I'm actually currently doing. I mostly use Codex for this as I want to save my Claude quota for the task at hand.
I built a version of this which wraps multiple CLI sessions locally. I do think the Web aspect and being able to access your CC session from anywhere is cool.
It really bothers me that it doesn't have support for devcontainers.
Only a closed set of languages are supported and the hook for startup installation of additional software seems to be not fully functioning at the moment.
You don't need claude code on the web for this, Cloudflare lets you spin up containers like crazy, you can boot an agent in a container, and as part of the boot process copy your claude auth token into the container. Then just ssh in, use tmux to make it persistent, and drive claude remotely.
Yeah and my preferred tools (mise) are missing from the environment, and installing it requires arcane environment configuration and then the LLM spends 10 minutes just trying to get the environment set up... On every interaction
Meanwhile, claude CLI has so many huge bugs that break the experience. Memory leaks, major cpu usage, tool call errors that require you to abandon a conversation, infinite loops, context leaks, flashing screens.. so many to list.
I love the feature set of Claude Code and my entire workflow has been fine tuned around it, but i had to to codex this month. Hopefully the Claude Code team spends some time to slow down and focus on bugs.
I doubt it. A large part of the performance problem with CC is constantly writing to a single shared JSON file across all instances, with no sharding or other mechanisms to keep it performant. It's spinning a shitload of CPU and blocking due to constant serialization/deserialization cycles and IO. When I was using CC a lot, my JSON file would hit >20mb quite quickly, and every instance would grind to a halt, sometimes taking >15s to respond to keyboard input. Seriously bullshit.
Everything Anthropic does from an engineering standpoint is bad, they're a decent research lab and that's it.
> Everything Anthropic does from an engineering standpoint is bad, they're a decent research lab and that's it.
This may be true, but then I wonder why it is still the case that no other agentic coding tool comes close to Claude Code.
Take Gemini Pro: excellent model let down by a horrible Gemini CLI. Why are the major AI companies not investing heavily in tooling? So far all the efforts I've seen from them are laughable. Every few weeks there is an announcement of a new tool, I go to try it, and soon drop it.
It seems to me that the current models are as good as they are goingto be for a long time, and a lot of the value to be had from LLMs going forward lies in the tooling
Gemini is a very powerful model, but it's tuned to be "oracular" rather than "agentic." The CLI isn't great but it's not the primary source of woe there. If you use Gemini with Aider in a more oracular fashion, it's still competitive with Claude using CC.
Claude is a very good model for "vibe coding" and content creation. It's got a highly collapsed distribution that causes it to produce good output with poor prompts. The problem is that collapsed distribution means it also tends to disobey more detailed prompts, and it also has a hard time with stuff that's slightly off manifold. Think of it like the car that test drives great but has no end of problems under atypical circumstances. It's also a naturally very agentic, autonomous model, so it does well in low information scenarios where it has to discover task details.
It is still slower than I'd like, at least with regards to UI input responsiveness, but I've never had it hard lock on me like CC. I can run 5-10 codex sessions and my system holds up fine (128GB RAM) but 8 CC instances would grind things to a halt after a few days of heavy usage.
I've been hoping that Claude Code on the Web also works with MCPs; so I can start getting it to do things beyond just coding. It's pretty awesome to use Git as a source of memory/tracking what's going on and pull requests as a way to build in a human-in-the-loop review flow.
> Git as a source of memory/tracking what's going on
That specific part doesn't have anything to do with Claude Web though, does it? When I use Codex and Claude they repeatedly look up stuff in the local git history when working on things I've mentioned I've worked on a branch or similar. As long as you make any sort of mention that you've used git, directly or indirectly, they'll go looking for it, is my feeling.
Love these discussions to find out what's new. For me replit.com is still the GOAT.
- Time to start your container (or past project) is ~1 sec to 1 min.
- Fully supported NixOS container with isolated, cloned agent layer. Most tools available locally to cut download times and ai web access risk.
- Github connections are persistent. Agents do a reasonable job with clean local commits.
- Very fast dev loops (plan/build/test/architect/fix/test/document/git commit / push to user layer) with adjustable user involvement.
- Phone app is fully featured... I've never built apps on roadtrips before replit.
- Uses claude code currently (has used chatgpt in the past).
Tips:
- Consider tig to help manage git from cli before you push to github.
- Gitlab can be connected but is clumsy with occasional server state refreshes.
- Startups that haven't committed to an IDE yet and expect compatibility with NixOS would have strong reason to consider this. It should save them the need to build their own OS-local AI code through early builds.
Honestly, I'm just flabbergasted at how incredible these tools are. I was able to build https://www.standup.net in a few days. Also was able update an old project https://www.microphonetest.com in a matter of hours with a plethora of features. Its truly addicting.
I was always disappointed by the Cursor version because the agents would make entirely new mistakes that Cursor IDE wouldn't make locally. Like so much that it was totally unusable. Completely messing up code edits to the point where a whole file would be deleted.
Interested to give this a go. But I would also need it to be able to run docker compose and playwright, to keep things on the rails.
In what area? I've been able to get it to do pretty much whatever I've tried it with so far, although probably Codex produces better code overall, even with the same prompts, and also have a web version. Although personally I prefer the CLIs.
Im still learning. All i know is claude.ai website chat. I thought claude code was a different thing. Not sure what codex is yet. Ive been using gemini assist in vscode for a week now, its kinda like just using it on the web but of course it edits your for you. Sometimes it ‘cant apply the changes though’
I have a question prompted by seeing what everyone is doing with Codex and Claude Code. I'm currently in a Data Analytics, B.S. program. I've thought of dropping out and focusing on coding with these AI tools, but some programmers have told me that by knowing SQL, Python, JavaScript and how to code in general, that it'll give me an advantage.
Is the 1.5 years that I have left worth it? (I already have an Associate's Degree).
I made an account just for you :) My opinion is: don't drop out.
1. The degree is useful. Having a Bachelor's opens up a lot of career paths because it shows that you committed to the Data Analytics program for four years. It also helps HR check off the "has a bachelor's" item on their list.
2. What you learn is useful. At the end of the day, you will be responsible for the code that the AI produces. How will you understand, explain, and justify your code to your colleagues and managers? "SQL, Python, JavaScript" and "theoretical Data Analytics knowledge" are both tools that will help you.
3. So far, senior engineers tend to have the most productivity boosts with AI. These engineers became "senior" before AI coding agents became mainstream, which means they know how to program. So based on this pattern, if you know how to program, then you will benefit more from AI.
Maybe you have other factors you are considering (e.g. money). My response is primarily based on the "existence of AI coding agents in the industry" factor.
I wish it didn't make public PRs to public repos. I sometimes fire off really speculative and sometimes silly requests and I really don't want a permanent record of these on an open source Github project. I could work on a fork but it's still fairly public.
Codex handles this much better. You choose when to make a PR and you can also just copy a .patch or git apply to your clipboard.
EDIT. They might have fixed this. Just testing. Does the mobile android app have Claude Code support yet or is it still annoyingly an iOS only thing?
EDIT2. It creates a public branch but not a PR. I'd still prefer that was a manual step.
I used Claude Code a lot until this weekend, when I gave Codex CLI a try, and I have to say, wow. The gpt-5-codex model is amazing. Sonnet 4.5 routinely gets stuff wrong, even Opus 4.1 isn't too amazing, but GPT 5 Codex just one-shots everything.
I've been using Sonnet whenever I run into the Codex limit, and the difference is stark. Twice yesterday I had to get Codex to fix something Sonnet just got entirely wrong.
I registered a domain a year ago (pine.town) and it came up for renewal, so I figured that, instead of deleting it, I'd build something on it, and came up with the idea of an infinite collaborative pixel canvas with a "cozy town" vibe. I have ZERO experience with frontend, yet Codex just built me the entire damn thing over two days of coding:
https://pine.town
It's the first model I can work with and be reasonably assured that the code won't go off the rails. I keep adding and adding code, and it hasn't become a mess of spaghetti yet. That having been said, I did catch Codex writing some backend code that could have been a few lines simpler, so I'm sure it's not as good as me at the stuff I know.
Then again, I wouldn't even have started this without Codex, so here we are.
It's interesting how different the subjective experiences of similarly-capable coding models is. My experience with Codex is that it tends to run off and do things without asking enough questions or keeping me in sync, whereas Claude seems to be more careful to clarify and keep me apprised of what it's doing.
I wonder how much of it comes down to how models "train us" to work in ways they are most effective.
I think a lot of it, Claude is definitely careful and Codex runs off too eagerly before discussing much (and the lack of a plan mode doesn't help), but I think we just learn how to use them. These days, anything I don't like goes into the AGENTS.md, where I tweak the instructions until the model understands well.
I too feel the eagerness, sometimes I only ask questions, it starts to code right away, and I have to add do not write any code very explicitly.
Codex has a plan mode.
In the web interface press the "+" button next to the repo it is working on. Not obvious at all though!
Sorry, I meant Codex CLI. It doesn't help that OpenAI has at least three things called Codex...
Weird, I tried the CoPilot and Codex CLIs and my experience was not good. I set it up with the same MCP tools I use elsewhere and the results were subpar compared to using agents in IDEs. I don't think it's a context issue either.
I’m not sure about op but codex high is slow but really solid. Med is definitely more hit or miss. I think the “meta” is claude does straightforward “shoe tying” code better and codex does more complicated thinking stuff better. Especially high.
Tool use codex is trash compared to sonnet. So still not a one stop shop.
This matches my experience. Codex is just more capable. But Sonnet is also quite good and much faster.
So lately I'll start with Sonnet for everything but the most complex tasks and then switch to Codex when needed.
> . I have ZERO experience with frontend,
After all these years, maybe even decades, of seeing your blog posts and projects on here, surely you must have had more experience with frontend than ZERO since you first appeared here? :)
Haha, fair, I meant "with React"!
He does have the experience... and stop calling me Shirley.
Any chance you could write up a blog post on your Codex experience(s)? Sounds really interesting.
I wrote a long comment but refreshed by accident before I could post it, so here we go again:
I'll write a post when I finish Pine Town, but I don't know what I could say about Codex in it. I think a big issue is that I don't know what others don't know, as the way I use LLMs (obviously) feels natural to me. Here are some tips that you may or may not already know:
* Reset the context as often as you can. LLMs like short contexts, so when you reach a point where the information has converged into something (e.g. the LLM has done a lot of work and you want it to change one of the details), reset the context, summarize what you want, and continue.
* Give the LLM small tasks that are logically coherent. Don't give it large, sprawling, open-ended tasks, but also don't give it chunks so tiny that it doesn't know what they're for.
* Explain the problem in detail, and don't dictate a solution. The LLM, like a person, needs to know why it's doing what it's doing, and maybe it can recommend better solutions.
* Ask it to challenge you. If you try to shoehorn the LLM too much, it might go off the rails trying to satisfy an impossible request. I've had a few times where it did crazy things because I didn't realize the thing I was asking for wasn't actually possible with the way the project was set up.
That's what I can think of off the top of my head, but maybe I'll write a general "how to work with LLMs" post. I don't think there's anything specifically different about Codex, and there must be a million such posts already, so I don't know if anyone will find value in the above... For me, it Just Worked™, but maybe that's just because I stumbled upon some specific technique that most people don't use.
Awesome! Looking forward to an in depth post once you get pinetown in order
my issue with codex is it will decide to take forever and do to much for one line changes I should've done myself, and sometimes would make more changes than desired. Claude Code is much more expedient and keeps its scope narrow and rarely goes outside the bounds of my request.
> sometimes would make more changes than desired
It's really easy to steer both Claude Code and Codex against that though, plop "Don't do any other changes than the ones requested" in the system prompt/AGENTS.md and they mostly do good with that.
I've tried the same with Gemini CLI and Gemini seems to mostly ignore the overall guidelines you setup for it, not sure why it's so much worse at that.
I agree with this, I've hit it too, plus I hit Codex limits in a day whereas I haven't hit a Claude limit yet, but all of this is more than compensated for by the simple fact that the code that Codex writes will almost always just work.
Sonnet is much less successful.
I am ride or die sonnet
Codex attempts to one shot for me but there’s many rounds of refinement. I haven’t used it in the last couple of weeks because it’s disappointing. Over hyped. Gone back to Amp and a little bit of Cursor with Sonnet 4.5
This is my entire problem with Codex - it will spend ten minutes trying to one shot a problem and usually go off the rails at some point, whereas Claude seems much better at incrementally finding the right solution with me.
I've heard this from many people, but I really haven't had this experience. Sonnet will write code that doesn't work, but Codex will give me working code basically every time. It does take longer, and it does think a lot, but I've never seen it go off the rails.
I do look at the backend code it writes, and it seems moderately sane. Sometimes it overcomplicates things, which makes me think that there are a few dragons in the frontend (I haven't looked), but by and large it's been ok.
> (I haven't looked)
Oh.
> I do look at the backend code it writes, and it seems moderately sane
Not good enough for you?
It's just a different way of approaching the problem, and might partially explain the preference for Codex' style.
If I'm doing a large task, I use GPT 5 Pro to write a spec first (with advice for Codex, broken down task list, snippets etc). I may also supply entire files/repos as context for 5 Pro to produce this.
If I skip 5 Pro but still have a large task, I have Codex write a spec file to use as a task list and to review for completeness as it works.
This is how you can use Codex without a plan mode.
I've noted it elsewhere, but Codex has a plan mode.
On the web, press the "+" button next to the repo
I still wish it would do all that on its own, without me having to switch models and make sure it won't make code changes.
Well, when you use GPT 5 Pro Mode it can't make any code changes, so not really a problem :)
I have similar workflow as parent, GPT 5 Pro for aiding with specifications and deep troubleshooting, rely on Codex to ground it in my actual code and project, and to execute the changes.
Codex won't read as much of your code as 5 Pro will (if you give it the context), and Codex will skip over reading in context that you give it (5 Pro can decide what's relevant after reading it all).
Yes Codex is still very early. We use it because it's the best model. The client experience will only get better from here. I noticed they onboarded a bunch of devs to the Codex project in GitHub around the time of 5's release.
> and Codex will skip over reading in context that you give it
That hasn't been my experience at all, neither first with the Codex UI since it was available to Pro users, nor since the CLI was available and I first started using that. GPT 5 Pro will (can, to be precise) only read what you give it, Codex goes out searching for what it needs, almost always.
That’s what I’m saying. Codex will search but then won’t read full files and is stingy with ingesting context. 5 Pro will take in a lot more context (quality up to about 60k input tokens) but you must give it. So sometimes you can even use Codex first to find what full files you should give to 5 Pro to create the spec/task list.
What my quote meant is that once you have the context Codex needs to do its work, if you give it to it, it’ll start the work right away without going and reading all those files again, which can help minimize context use within a Codex session (by having 5 Pro or just another Codex read in a lot of context to identify what is relevant for Codex instead of having Codex waste precious context headroom on discovery in a session that is dedicated to doing the work).
Yeah I was using Claude pretty continuously for 3, 4 months and then decided to give Codex a whirl and it was impressive. I'd consider it to be a lot more cautious and careful and less lazy?
It is however slow, and more expensive. You can either pay the $20 and get maybe 2 days of work out of it, or $200 for "Pro." But there's nothing inbetween like the $100 USD Claude Code tier.
Yeah, I'm really missing the $100 tier. The $20 gets me a day of coding a week with it, which is way too little, and $200/mo is too much for hobby projects.
I've personally been running the Claude Code tool but pointed at DeepSeek's API platform. Cheaper than both Anthropic and OpenAI, and about as good as Sonnet 4 was, I'm finding.
Context window is too small though, and it sometimes has problems with compacting. But I was having that with Sonnet 4.5 as well.
I really loved using Claude. I like working with Claude more than GPT or Gemini. Claude is to LLMs what Firefox is to browsers. I just like Firefox more than Chrome. It's very clearly behind GPT Codex at this point though. So far I've found Gemini for front-end design work to be better than the others, and I pair it with GPT for everything else. Hopefully Gemini 3 is a solid improvement, I like having at least two LLMs at high quality to run against each other.
Claude Code is much better than Codex CLI, but GPT 5 Codex is much better than Sonnet 4.5. I wish I could use one with the other, but alas.
There are tools like claude-code-router. I've gone through the pain of getting gpt-5, gemini-2.5-pro, and other models wired together. The system prompt differences are too much though I think, claude still feels the best in claude code.
I'm at the point where I have so much built up around claude code workflows that claude feels very good. But when I don't use them, I find that I immensely prefer gpt-5 (and for harder, design influencing questions, grok-4 heavy which is not available behind an API)
Yeah, I think the system prompts are so optimised for the specific model that others won't work as well, so it kind of defeats the purpose of being able to plug your own model in. I wish I could, but I know I won't get as good performance as with the model's native cli.
huh, my current ranking is:
1. claude code CLI, generally works, great tool use
2. codex on the web, feels REALLY smart, but can’t use tools
3. codex CLI, still smarter than claude but less situational awareness
4. codex via iphone app, buggier than the web app
5. claude code on the web, worst of all worlds
I like the workflow with Codex more. Though I like working with Claude more. So I wish Anthropic would copy the Codex workflow.
I like that Codex commits using your identity as if it was your changes. And I like that you can interact with it directly from the PR as if it was a team member.
You can instruct Claude Code to commit in your name. Tell it in the CLAUDE.md file. Or add via `# Commit as xyz` and it will memorize.
Also add `"includeCoAuthoredBy": false` to your `settings.json` file (you may also need to reinforce this in your commit prompt YMMV).
ahhhhh thank you! this saves me from having to add this to every repo's CLAUDE.md file.
Ah, excellent. Thanks for sharing.
The whole flow of:
creating container -> cloning repo -> making change -> test -> send PR
is too slow of a loop for me to do anything much useful. It's only good for trivial "one-shot" stuff.
I'd say this method of coding agent interaction is likely a strong contender for integrating coding agents into teams. You start with a really well defined ticket and a good source of relevant documentation for the project then set the agent loose by assigning it a ticket. It does it's thing, maybe asks questions on a group chat or in the ticket, and eventually produces a PR for the ticket. It's the 'interface' behind how a developer interacts with a project already. There's a lot of hand-waving in there and it's not a today or tomorrow thing, but it seems like it's coming fairly soon.
OpenHands does that, I wrote about it if you search my submissions.
thats the premise behind the popular Devin. I don't think it saw any market fit.
I use it (and Codex web) specifically when I'm not at my desk (or I am but in the middle of something else) and I want to do something fairly speculative. Kinda either exploratory or investigative. I may or may not use the results but it doesn't get in the way of anything I'm actually currently doing. I mostly use Codex for this as I want to save my Claude quota for the task at hand.
I built a version of this which wraps multiple CLI sessions locally. I do think the Web aspect and being able to access your CC session from anywhere is cool.
https://github.com/built-by-as/FleetCode
It really bothers me that it doesn't have support for devcontainers.
Only a closed set of languages are supported and the hook for startup installation of additional software seems to be not fully functioning at the moment.
You don't need claude code on the web for this, Cloudflare lets you spin up containers like crazy, you can boot an agent in a container, and as part of the boot process copy your claude auth token into the container. Then just ssh in, use tmux to make it persistent, and drive claude remotely.
Yeah and my preferred tools (mise) are missing from the environment, and installing it requires arcane environment configuration and then the LLM spends 10 minutes just trying to get the environment set up... On every interaction
Meanwhile, claude CLI has so many huge bugs that break the experience. Memory leaks, major cpu usage, tool call errors that require you to abandon a conversation, infinite loops, context leaks, flashing screens.. so many to list.
I love the feature set of Claude Code and my entire workflow has been fine tuned around it, but i had to to codex this month. Hopefully the Claude Code team spends some time to slow down and focus on bugs.
I doubt it. A large part of the performance problem with CC is constantly writing to a single shared JSON file across all instances, with no sharding or other mechanisms to keep it performant. It's spinning a shitload of CPU and blocking due to constant serialization/deserialization cycles and IO. When I was using CC a lot, my JSON file would hit >20mb quite quickly, and every instance would grind to a halt, sometimes taking >15s to respond to keyboard input. Seriously bullshit.
Everything Anthropic does from an engineering standpoint is bad, they're a decent research lab and that's it.
> Everything Anthropic does from an engineering standpoint is bad, they're a decent research lab and that's it.
This may be true, but then I wonder why it is still the case that no other agentic coding tool comes close to Claude Code.
Take Gemini Pro: excellent model let down by a horrible Gemini CLI. Why are the major AI companies not investing heavily in tooling? So far all the efforts I've seen from them are laughable. Every few weeks there is an announcement of a new tool, I go to try it, and soon drop it.
It seems to me that the current models are as good as they are goingto be for a long time, and a lot of the value to be had from LLMs going forward lies in the tooling
Gemini is a very powerful model, but it's tuned to be "oracular" rather than "agentic." The CLI isn't great but it's not the primary source of woe there. If you use Gemini with Aider in a more oracular fashion, it's still competitive with Claude using CC.
Claude is a very good model for "vibe coding" and content creation. It's got a highly collapsed distribution that causes it to produce good output with poor prompts. The problem is that collapsed distribution means it also tends to disobey more detailed prompts, and it also has a hard time with stuff that's slightly off manifold. Think of it like the car that test drives great but has no end of problems under atypical circumstances. It's also a naturally very agentic, autonomous model, so it does well in low information scenarios where it has to discover task details.
Is codex cli performant? I've been on codex all month and it seems to chew through my battery just like claude code did.
It is still slower than I'd like, at least with regards to UI input responsiveness, but I've never had it hard lock on me like CC. I can run 5-10 codex sessions and my system holds up fine (128GB RAM) but 8 CC instances would grind things to a halt after a few days of heavy usage.
Ah, yeah - same for me on that front.
Just showing a question causes CC to spin a cpu core at 100%.
I would love for them to open up the API to this.
I'd like to build an integration with Whisper Memos (https://whispermemos.com/)
Then I'd be able to dictate a note on my Apple Watch such as:
> Go into repository X and look at the screen Y, and fix bug Z.
That'd be so cool.
I've been hoping that Claude Code on the Web also works with MCPs; so I can start getting it to do things beyond just coding. It's pretty awesome to use Git as a source of memory/tracking what's going on and pull requests as a way to build in a human-in-the-loop review flow.
> Git as a source of memory/tracking what's going on
That specific part doesn't have anything to do with Claude Web though, does it? When I use Codex and Claude they repeatedly look up stuff in the local git history when working on things I've mentioned I've worked on a branch or similar. As long as you make any sort of mention that you've used git, directly or indirectly, they'll go looking for it, is my feeling.
Love these discussions to find out what's new. For me replit.com is still the GOAT.
- Time to start your container (or past project) is ~1 sec to 1 min. - Fully supported NixOS container with isolated, cloned agent layer. Most tools available locally to cut download times and ai web access risk. - Github connections are persistent. Agents do a reasonable job with clean local commits. - Very fast dev loops (plan/build/test/architect/fix/test/document/git commit / push to user layer) with adjustable user involvement. - Phone app is fully featured... I've never built apps on roadtrips before replit. - Uses claude code currently (has used chatgpt in the past).
Tips: - Consider tig to help manage git from cli before you push to github. - Gitlab can be connected but is clumsy with occasional server state refreshes. - Startups that haven't committed to an IDE yet and expect compatibility with NixOS would have strong reason to consider this. It should save them the need to build their own OS-local AI code through early builds.
What does this mean for products like Terragon and Sculptor?
Agreed. I can vibe code from an iPad now. Workflow is Claude Code for Web + Vercel.
Honestly, I'm just flabbergasted at how incredible these tools are. I was able to build https://www.standup.net in a few days. Also was able update an old project https://www.microphonetest.com in a matter of hours with a plethora of features. Its truly addicting.
I was always disappointed by the Cursor version because the agents would make entirely new mistakes that Cursor IDE wouldn't make locally. Like so much that it was totally unusable. Completely messing up code edits to the point where a whole file would be deleted.
Interested to give this a go. But I would also need it to be able to run docker compose and playwright, to keep things on the rails.
Any good demos of what claude code can do?
In what area? I've been able to get it to do pretty much whatever I've tried it with so far, although probably Codex produces better code overall, even with the same prompts, and also have a web version. Although personally I prefer the CLIs.
Im still learning. All i know is claude.ai website chat. I thought claude code was a different thing. Not sure what codex is yet. Ive been using gemini assist in vscode for a week now, its kinda like just using it on the web but of course it edits your for you. Sometimes it ‘cant apply the changes though’
I have a question prompted by seeing what everyone is doing with Codex and Claude Code. I'm currently in a Data Analytics, B.S. program. I've thought of dropping out and focusing on coding with these AI tools, but some programmers have told me that by knowing SQL, Python, JavaScript and how to code in general, that it'll give me an advantage.
Is the 1.5 years that I have left worth it? (I already have an Associate's Degree).
I made an account just for you :) My opinion is: don't drop out.
1. The degree is useful. Having a Bachelor's opens up a lot of career paths because it shows that you committed to the Data Analytics program for four years. It also helps HR check off the "has a bachelor's" item on their list.
2. What you learn is useful. At the end of the day, you will be responsible for the code that the AI produces. How will you understand, explain, and justify your code to your colleagues and managers? "SQL, Python, JavaScript" and "theoretical Data Analytics knowledge" are both tools that will help you.
3. So far, senior engineers tend to have the most productivity boosts with AI. These engineers became "senior" before AI coding agents became mainstream, which means they know how to program. So based on this pattern, if you know how to program, then you will benefit more from AI.
Maybe you have other factors you are considering (e.g. money). My response is primarily based on the "existence of AI coding agents in the industry" factor.
We no longer swoon over IDE features but now Llm correctness and novelty.
IDEs were good tools for humans to write code in, but are no longer needed.
Likely not true.
Citation needed.
I wish it didn't make public PRs to public repos. I sometimes fire off really speculative and sometimes silly requests and I really don't want a permanent record of these on an open source Github project. I could work on a fork but it's still fairly public.
Codex handles this much better. You choose when to make a PR and you can also just copy a .patch or git apply to your clipboard.
EDIT. They might have fixed this. Just testing. Does the mobile android app have Claude Code support yet or is it still annoyingly an iOS only thing?
EDIT2. It creates a public branch but not a PR. I'd still prefer that was a manual step.
How would it push stuff to a public GH repo without the pushed commits being public? This seems like a GitHub limitation, rather than a Claude one.
Don't push at all until I authorize it. That's what Codex does.
The web app? How do you look at the code it wrote? I've only used the cli.