We’re releasing GPT-5-Codex—a version of GPT-5 further optimized for agentic coding in Codex.
It’s now live everywhere you use Codex—in your terminal or IDE, on the web, in GitHub, and in the ChatGPT iOS app.
We’re working on optimizing GPT-5-Codex for use in other agentic coding tools, and will bring it to the API soon.
More in our blog: https://openai.com/index/introducing-upgrades-to-codex/
Thanks for the update, but GPT-5 Codex currently has a critical bug: execution speed is extremely slow even on very simple tasks, and if it “thinks” for too long it just crashes with an error and no result. This makes it unusable for real coding workflows. Please investigate urgently.
Same problem here. GPT5 Codex is very very slow. Unusable. Simple tasks that it used to complete in seconds now take up to 20 minutes. It’s truly unusable.
Yeah, complex tasks failing with “Failed to sample tokens.”.
Everything working well.
Thank you for fixing!
I will keep using claude code until you decide to offer flat plans !
Same here as well, running on a 10000 line code it is taking 15 minuts sometimes to find one line to change - make it somewhat unusable
Welcome to the community Max Zhadobin!
The usage limits on plus are crazy… I cant afford a pro plan, however on 2 plus plans i could do 5 days of work, and then both would be done. Since the new codex… Which yes, is painfully slow, I have hit both caps in 1 day and a few hours (NOT RUN SIMULTANEOUSLY). Its not 2 on one account, they are separate so should each get their own usage cap? So How the F??? I have done 5 times the work previously in the same amount of time, and now I must wait almost 6 days to work again. THIS IS LUDICROUS!
I have been waiting all week for an API updated to make this available for custom agents.
openai.NotFoundError: Error code: 404 - {'error': {'message': 'The model `gpt-5-codex` does not exist or you do not have access to it.', 'type': 'invalid_request_error', 'param': None, 'code': 'model_not_found'}}
I hope that this is added soon.
I gave it a try in vscode. It is a dumb model. A half dozen “searching” code with powershell and getting nowhere, when what it needed was entire file loaded to context for observation, and I’m done.
Seconded because it gpt-5-codex doesn’t work with Codex CLI if using an API key. (Our company doesn’t issue ChatGPT accounts, but does let us use API keys charged to the corporate account.)
Ideas of when we can get this support?
I support the attempt to make GPT coexist with other IDEs.
For now, even the low-end model in Windsurf seems to perform well enough, but I feel uneasy that I cannot see how credits are being consumed.
At present, I’ve noticed the following behaviors:
It inserts carriage return codes (\r\n) into the source code.
It keeps wandering around the file and never returns.
Refactoring sometimes fails and causes methods to escape from their classes.
Multibyte characters get corrupted.
I would like these issues—especially the ones that cannot be avoided by user instructions—to be prioritized and fixed first.
The problem of excessively long search times might be avoided by narrowing the scope of topics in a prompt (session) and by making use of the AGENTS.md file and the Auto Context button.
As for runtime logs, I think they should show only proof that the system is following the user’s instructions, rather than dumping raw computer activity. A huge list of PowerShell commands may be scientifically accurate, but it is hard to read.
I’m not sure if something changed when the new model came out but I have noticed a significant increase in the rate limits (Pro).
1-2 weeks ago, I was hitting the weekly capacity every single day. And it did not take a full week, I’ll tell you that.
Now I am raw dogging 6+ consoles all day and night on high – no limits (yet). Big improvement.
THX SAM!
Edit: Just seeing some of the above comments about gpt-5-codex … I don’t want to be That Guy, but I’ve consumed tens of billions of tokens this year… and it is the best agentic coding model in the world right now, bar none. It’s really remarkable. It’s so far and above all other CLI models including Claude Code (which isn’t saying much nowadays) that I am surprised there has not been much more fanfare over it even compared to gpt-5-high! If I had to pick one qualm, it’s the degradation with context growth. The speed is actually great.
Again, THX SAM.
I have been extremely happy with gpt 5 codex high in codex cli. I have not hit rate limits too often, and it gets the job done nicely. I love that it does not go overboard with comments like so many models tend to. Overall this is tremendous value for 20-30 dollars a month
@edwinarbus any updates on when gpt5-codex will be available through API. thanks
I was hit by a rate limit after less than 36 hours of subscribing, doing some light work, and on the first day, extremely slow work. Locked me out for 5 1/2 days. Unacceptable.
kinda funny to read messages about someone hitting rate limits after few days of work but not being able to afford pro accoumt 
guys are u like working for free or u don’t understand how much is to run service like this and u want it for 20usd? nevermind…
but more importantly, when gpt-5-codex on API?
i am tier 5 and would love to use that with API already.
also having little issues with speed of gpt-5 high , obviously you need to throttle because of high load, so sometimes i use service_tier=priority , still worth it…
i can lend u my 12 x rtx 3090 server if that helps 
just please activate gpt-5-codex on API already