people are sleeping on openai right now but codex 5.2 xhigh is at least as good as opus and you get a TON more usage out of the OpenAI $20/mo plan than Claude's $20/mo plan. I'm always hitting the 5 hour quota with Opus but never have with Codex. Codex tool itself is not quite as good but close.
I dont think this is Cerebras. Running on cerebras would change model behavior a bit and it could potentially get a ~10x speedup and it'd be more expensive. So most likely this is them writing new more optimized kernels for Blackwell series maybe?
OpenAI in my estimation has the habit of dropping a model's quality after its introduction. I definitely recall the web ChatGPT 5.2 being a lot better when it was introduced. A week or two later, its quality suddenly dropped. The initial high looked to be to throw off journalists and benchmarks. As such, nothing that OpenAI says in terms of model speed can be trusted. All they have to do is lower the reasoning effort on average, and boom, it becomes 40% faster. I hope I am wrong, because if I am right, it's a con game.
Starting off the ChatGPT Plus web users with the Pro model, then later swapping it for the Standard model -- would meet the claims of model behavior consistency, while still qualifying as shenanigans.
It's good to be skeptical, but I'm happy to share that we don't pull shenanigans like this. We actually take quite a bit of care to report evals fairly, keep API model behavior constant, and track down reports of degraded performance in case we've accidentally introduced bugs. If we were degrading model behavior, it would be pretty easy to catch us with evals against our API.
In this particular case, I'm happy to report that the speedup is time per token, so it's not a gimmick from outputting fewer tokens at lower reasoning effort. Model weights and quality remain the same.
I've seen Sam Altman make similar claims in interviews, and I now interpret every statement from an Open AI employee (and especially Sam) as if an Aes Sedai had said it.
I.e.: "keep API model behavior constant" says nothing about the consumer ChatGPT web app, mobile apps, third-party integrations, etc.
Similarly, it might mean very specifically that a "certain model timestamp" remains constant but the generic "-latest" or whatever model name auto-updates "for your convenience" to the new faster performance achieved through quantisation or reduced thinking time.
You might be telling the full, unvarnished truth, but after many similar claims from OpenAI that turned out to be only technically true, I remain sceptical.
It looks like you do pull shenanigans like these [0]. The person you're replying to even mentioned "ChatGPT 5.2", but you're specifically talking only about the API, while making it sound like it applies across the board. Also appreciate the attempt to further hide this degradation of the product they paid for from users by blocking the prompt used to figure this out.
Hey Ted, can you confirm whether this 40% improvement is specific to API customers or if that's just a wording thing because this is the OpenAI Developers account posting?
its worth you guys doing on your end, some analysis of why customers are getting worse results a week or two later, and putting out some guidelines about what context is poisonous and the like
Starting off the ChatGPT Plus web users with the Pro model, then later swapping it for the Standard model -- would meet the claims of model behavior consistency, while still qualifying as shenanigans.
It was probably from the other day when roon realized that normal people have it slower than staff.
Then from that they realized they could just run API calls more like staff, fast, not at capacity.
Then they leave the billion other people's calls at remaining capacity.
https://thezvi.substack.com/i/185423735/choose-your-fighter
> Ohqay: Do you get faster speeds on your work account?
> roon: yea it’s super fast bc im sure we’re not running internal deployment at full load
This is great.
In the past month, OpenAI has released for codex users:
- subagents support
- a better multi agent interface (codex app)
- 40% faster inference
No joke, with the first two my productivity is already up like 3x. I am so stoked to try this out.
this is for api only
Shoot me
looks like i'm wrong
Try Claude and you can get x^2 performance. OpenAI is sweating
May be a bit different depending on what kind of work you're doing, but for me 5.2-codex finally reached higher level than opus.
5.2-codex is pretty solid and you get dramatically higher usage rates with cheap plans. I would assume API use is much cheaper as well.
people are sleeping on openai right now but codex 5.2 xhigh is at least as good as opus and you get a TON more usage out of the OpenAI $20/mo plan than Claude's $20/mo plan. I'm always hitting the 5 hour quota with Opus but never have with Codex. Codex tool itself is not quite as good but close.
It’s interesting that they kept the price the same while doing inference on Cerebras is much more expensive.
I dont think this is Cerebras. Running on cerebras would change model behavior a bit and it could potentially get a ~10x speedup and it'd be more expensive. So most likely this is them writing new more optimized kernels for Blackwell series maybe?
Fair point but it remains to answer - why isn’t this speed up available in ChatGPT and only in the api?
this is almost certainly not being done on cerebras
OpenAI in my estimation has the habit of dropping a model's quality after its introduction. I definitely recall the web ChatGPT 5.2 being a lot better when it was introduced. A week or two later, its quality suddenly dropped. The initial high looked to be to throw off journalists and benchmarks. As such, nothing that OpenAI says in terms of model speed can be trusted. All they have to do is lower the reasoning effort on average, and boom, it becomes 40% faster. I hope I am wrong, because if I am right, it's a con game.
Starting off the ChatGPT Plus web users with the Pro model, then later swapping it for the Standard model -- would meet the claims of model behavior consistency, while still qualifying as shenanigans.
It's good to be skeptical, but I'm happy to share that we don't pull shenanigans like this. We actually take quite a bit of care to report evals fairly, keep API model behavior constant, and track down reports of degraded performance in case we've accidentally introduced bugs. If we were degrading model behavior, it would be pretty easy to catch us with evals against our API.
In this particular case, I'm happy to report that the speedup is time per token, so it's not a gimmick from outputting fewer tokens at lower reasoning effort. Model weights and quality remain the same.
I've seen Sam Altman make similar claims in interviews, and I now interpret every statement from an Open AI employee (and especially Sam) as if an Aes Sedai had said it.
I.e.: "keep API model behavior constant" says nothing about the consumer ChatGPT web app, mobile apps, third-party integrations, etc.
Similarly, it might mean very specifically that a "certain model timestamp" remains constant but the generic "-latest" or whatever model name auto-updates "for your convenience" to the new faster performance achieved through quantisation or reduced thinking time.
You might be telling the full, unvarnished truth, but after many similar claims from OpenAI that turned out to be only technically true, I remain sceptical.
It looks like you do pull shenanigans like these [0]. The person you're replying to even mentioned "ChatGPT 5.2", but you're specifically talking only about the API, while making it sound like it applies across the board. Also appreciate the attempt to further hide this degradation of the product they paid for from users by blocking the prompt used to figure this out.
Happy to retract if you can state [0] is false.
[0] https://x.com/btibor91/status/2018754586123890717
Hey Ted, can you confirm whether this 40% improvement is specific to API customers or if that's just a wording thing because this is the OpenAI Developers account posting?
so what actually happens if it isnt shenanigans?
its worth you guys doing on your end, some analysis of why customers are getting worse results a week or two later, and putting out some guidelines about what context is poisonous and the like
Starting off the ChatGPT Plus web users with the Pro model, then later swapping it for the Standard model -- would meet the claims of model behavior consistency, while still qualifying as shenanigans.
You're confirming you don't alter "juice" levels..?
I mean you can just run the benchmark again
tons of posts on reddit that they also significantly dropped quality