severe degradation in Codex CLI
What version of Codex is running?
codex-cli 0.45.0
Which model were you using?
gpt-5-codex
What platform is your computer?
Linux 6.6.87.2-microsoft-standard-WSL2 x86_64 x86_64
What steps can reproduce the bug?
Ask Codex to implement a small/medium size feature (300-800 LoC) within an existing framework.
Let Codex go forward with development.
Branch A:
Report an issue with and/or request a change to its approach, or otherwise tell it to change the architecture used.
Observe Codex ignores instructions, repeats failures, and continues to pursue the "original" implementation. Continued attempts to correct results in greater deviations from any established requirements or instruction.
Tell it to start over from scratch, observe it deletes some code and then reimplements the same approach.
Give up and implement yourself.
Branch B:
Observe trainwreck of an implementation and just give up right away.
Implement yourself, saving considerable time and mental anguish resulting from hours of fruitless back-and-forth.
What is the expected behavior?
Quality of solutions produced by Codex does not suddenly degrade one day to the next.
What do you see instead?
Unusable code and a wasted day.
Additional information
It's the same behavior Claude Code displayed about a month ago before losing numerous users that were consistently reporting sudden performance/quality issues.
Up to this point I've had a (more or less) perfectly functional relationship with Codex CLI and models. But today, between both my personal Pro license and my work's Enterprise license, I witnessed incredible failures to adhere to instructions or produce usable results. Two separate and wholly different projects, identical behaviors.
I'd be interested to hear if OpenAI is aware of an issue with the service?
5 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Reviewed, no duplicates detected.
_Powered by Tired Human_
Maybe it's because there are too many people using it. A lot of CC second-hand transfer number dealers have gone to Codex.
You might be right about high-traffic times causing issues. A couple hours after I made the original post it felt like it was "back to normal."
Today's daytime experience mirrors that of the original post. Codex nearly destroyed my project in the background while I was squinting and straining what's left of my brain to understand how it made the leap on some initial code changes it decided to make. I should've cancelled the run immediately. Fortunately the project is on a VCS and I've "banned" it from using git commands, or it likely would've just wiped my local branch for fun. I'm glad it managed to follow that restriction at least. Still, that'll teach me to leave approvals on auto.
I'd be curious to understand how increased load results in reduced quality. On the surface I wouldn't expect anything to change unless OpenAI is automatically scaling down to lower quants to deal with increased traffic without scaling infrastructure; the only thing I might expect is increased roundtrip time while the API queues individual requests, but no loss in quality of the response itself. At least that's what I, and the business I work for, are paying for... consistent quality.
Thanks for the bug report. We've shared the feedback with the team members who are responsible for training our codex models. We're tracking this feedback and similar reports so we can make the model more robust in future iterations.
If you see other model behavior that you'd like to report, please use the recently-added
/feedbackcommand in the CLI. That will provide us with additional details including logs and conversation text.