`GPT-5.2-High` feels off

Resolved 💬 9 comments Opened Jan 28, 2026 by guidedways Closed Jan 29, 2026

💡 Likely answer: A maintainer (guidedways, contributor) responded on this thread — see the highlighted reply below.

What version of Codex is running?

0.92.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.2-high

What platform is your computer?

macOS

What terminal emulator and version are you using (if applicable)?

iTerm

What issue are you seeing?

I'm either mistaken or there's something off about GPT 5.2-High and even Xhigh (non-codex models) after the Rate-Limit warnings showed up this morning. It feels as if under the hood it's the 'codex' version at work - it's no longer behaving as consistently as it has for the past 1 month. Noticeably, it feels 'lazy' - the hallmark of gpt-5.2-codex-high when it comes to implementation.

For instance, I explicitly instructed that it make some changes and then confirm that the app builds. It did that once but later down the line it's making edits and no longer bothers to check if it left the code in an uncompilable state - gpt-5.2-high has never done anything like this, but the codex models instead always have (in my experience).

Am I seeing things?

What steps can reproduce the bug?

Difficult to quantify in steps.

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

9 Comments

matssk · 5 months ago

I can say that I did wonder about the exact same thing. It seemingly has to be pushed more to finish tasks, i.e. ask "did you do all the tasks in the plan?", but it's hard to say for sure.

guidedways contributor · 5 months ago

Yes that is it - I've switched to gpt-5.2-xhigh and I _think_ it's performing better than high. I'm a little unsure what's going on, just feels a little different today.

Dr-Aniekan-Udo · 5 months ago

I have the same issue with the gpt-5.2-high after the rate-limiting warning early this morning. It can't even follow instructions again, and I have been working with it successfully until this morning.

matssk · 5 months ago

So right now I have a session where it's going on and on again as usual, not being lazy at all. Earlier today and yesterday I've had several session where it was being lazy and and not really finishing tasks, so I don't know if this was a temporary issue? Or just getting unlucky with some sessions perhaps?

guidedways contributor · 5 months ago

I'm on xhigh now and it's doing what it usually does, scared to switch back to high 😅

AyinMostima · 5 months ago

So right now I have a session where it's going on and on again as usual, not being lazy at all. Earlier today and yesterday I've had several session where it was being lazy and and not really finishing tasks, so I don't know if this was a temporary issue? Or just getting unlucky with some sessions perhaps?

I also ran into this issue. Codex is currently very unstable, but they definitely won't give a direct answer.

etraut-openai contributor · 5 months ago

We're not aware of any backend changes that would explain the observations in this thread. The rate limiting issue that some users experienced yesterday was not at all related to model performance or output.

guidedways contributor · 5 months ago

@etraut-openai appreciate the reassurance. We probably are over analyzing this. xhigh at least has been working just as well as it always has.

Xynonners · 5 months ago

I can say that I did wonder about the exact same thing. It seemingly has to be pushed more to finish tasks, i.e. ask "did you do all the tasks in the plan?", but it's hard to say for sure.

for some reason, in the reasoning summaries sometimes it talks as if it's running out of time or time is limited and takes shortcuts as a result.