GPT 5.5 seems to be degraded

Codex Codex CLI
32

Just reporting back here, I cannot tell if the issue has been fixed but it is catching the errors that were caused before the fix, so I don’t know if that is an indication of improvement or not.

What does surprise me is that this is 5.5-xhigh and its clear that I have to use the most expensive model to really get meaningful work done. Previously, before the degradations were noticed, I wouldn’t have reached xhigh except for planning or one off problem solving that 5.5-high or med could not fix.

I don’t have enough information or confidence yet to switch back to using gpt-5.5-med/high yet. I will continue trying different prompts to see if it can yield any more meaningful changes.

Thanks again for your attention to this matter.

33

I’m reporting back again and I definitely don’t think the issue has been fixed. I can’t quite put a finger on it but there’s just something about it that feels off still. Normally, I could always rely on xhigh to break through a tough problem but it still doesn’t seem as capable.

Basically I’m working with Flutter and lot of dart code and I have been using codex since September last year and at its current state, its quite frustrating because it feels like we’ve regressed.

Also this shouldn’t be a context issue as I’m running gpt 5.5 xhigh on a fairly new convo. I wish I could offer more information but its tough to convey this “vibe” so to speak, all I can tell you is that it does not seem to have the same punch as it used to. I don’t think the codebase has grown large either, its just since a few weeks ago I started noticing gpt 5.5 make more sloppy mistakes and I don’t think we are quite out of the forest yet.

I’ve finally had to resort to Opus 4.8 to audit gpt-5.5-xhigh’s work and it confirmed the very crux of my complaint with 5.5 that it did surface level edits dressed up as deep work and it escalated a lot of false positives as legitimate problems to solve without consulting me.

Once Opus 4.8 was identify my intent by analyzing GPT 5.5’s code, it was able to one shot all the issues that 5.5-xhigh was “working on” past few days…

Again its of subjective nature but Opus 4.8 confirms the common critique raised of GPT 5.5 and thus raises another concerning point that GPT-5.5-xhigh was not able to catch itself and dig us out of the hole.

This isn’t the first time I’ve experienced this (saw it in 5.4 too) and originally I was very excited to use GPT 5.5 as it appeared to have addressed the shortcomings but it seems like this tendency to perform “exhibitions” rather than truly deep work and not recognizing issue is back.

This is the best I can do in offering an insight to what feels “off” I hope it is helpful in improving the model but I do wonder why we constantly go through this roller coaster ride where a model feels great and then right around the time a new model is expected things just seem less put together.

I’ve not used Claude in months since codex has taken over ($200/month plan here) but this session with Opus 4.8 crystalizes the issues I’ve been having with 5.5 recently and first time I am asking whether I should invest more into Claude. I know Opus 4.8 just came out and I need to run more tests but a $20/month plan suddenly solving what a $200/month plan has been struggling with past few days makes its value proposition worth while to explore integrating into my workflow.

I know the OpenAI team is hard at work here and I trust that they can make this work!

Thank you again for reading this long rant/review.

34

This absolutely is the case. I just renewed subscription and honestly I regret about it.

Instruction following has became just terrible.

E.g. mentioning harsh ETA for a feature long ago in the chat history somehow made model extrapolate it to the whole work, so it reflected upon itself, saying it implied it had been told to “rush things” (it never had been)

It became viscous - it doesn’t prioritize the current focus and direct steering over the rest of the context. This is clear regression and it’s making it nearly impossible to deliver something of proper quality.

Worth nothing that code quality itself has terribly degraded as well.

Being communicated clear strict quality gates using native language features and align to the language BPs (in my case - nix), it just created a long messy bash script inside a nix derivation callback.

No changes were made in configuration on my end to blame.
No style changes in how I steer it.

Please do something about it, OpenAI.

35

I am also facing the same issue, it feels like using gpt 5.3

36

Thanks for flagging this, everyone. I want to echo @Prashant_Pardesi's suggestion.

If you're seeing this issue, please send a report to support@openai.com after doing the following:

  1. Reproduce the issue in a new Codex chat.
  2. Right-click and copy the Session ID from the affected chat.
  3. Use the /feedback command and include as much detail as possible about what you're experiencing.
  4. Share the Session ID and any relevant screenshots or examples in your support request.

Once you've submitted the ticket, please post the case ID here. That will help the team track reports and investigate the issue more efficiently.

If anyone notices specific prompts or workflows that consistently trigger the problem, feel free to share those details as well.

-Mark G.

37

The problems with image creation prompts are insane. It randomly becomes very stupid. I need to give the image reference everytime as it does not use the same image even I edit itself.

38

I’m having the same issues as everybody else. This is very disturbing going backwards to 5.4. I’m getting tired of being a beta test site and not getting paid for it or some kind of discount for staying as a loyal customer. When this subscription is up, I think I’m moving to Claude. Opus 4.8 is looking good.

39

Same here. I asked it to create an SVG file, but it never got it correct. However, Opus 4.8 just one-shotted it to finish.

40

Degraded is the understatement of the century.

  • It suddenly literally crawling. Runs that used to take a couple of minutes take over 30. On very simple stuff.

  • It has started severely ignoring rules, and repeating things that it was told not to ever do again.

  • It tries to do things from earlier context that are completely irrelevant.

  • It tries to cut corners. 5.5 high, example. I give it a very specific direction on what to do in C++ code (and I don’t mean vibe coding prompt, I mean a detailed engineering plan with specific patterns and even code examples) and it finds a static helper that by name sounds similar to what I asked it to write, and without examining that static helper, its references and what it actually does, decides to ignore my direction and alter that static helper in a way that breaks everything. It has clear directions to do exactly what I say and nothing else. Not to deviate at all, not to bias on minimal change, not to try hackish, brittle or random timer based stuff (oh boy does GPT love setting up random timers with arbitrary timeouts in order to deal with deferrals/queuing instead of using proper event/signal/etc patterns).

    5.5 high used to do nothing of the above. This feels like 3.x era GPT.

41

I think your C++ code is very close to what I’ve discovered is that it simply does only what it thinks is correct using the minimum context possible without fully understanding the larger picture of what its doing. That wasn’t the case before which is why I’m surprised this keeps happening. It’s not like it can’t do a good job either, I’m beginning to think this is the limit of the context size

In particular are tests it has written which are literally just a bunch of regex matches.

I am now using Opus 4.8 to drive development and have it tell GPT 5.5 what to do. I would never imagine that I need to do this with GPT 5.5 xhigh.

Really hope they address this problem soon I am considering downgrading my codex to $100/month . GPT 5.5 is stiill full of potential it has performed great to date until the degradation.

42

Yes, it looks like aggressive optimizations to reduce context size/take the shortest path possible. Or something like that. Still doesn’t exactly explain the blowing up of time. Or how suddenly it started doing broad replace patches with minimal replace patterns that touch unrelated parts of the code, it sees that, it tries again for 3-4 times before switching to minimal set accurate patches, corrects the bad unintended changes, then immediately tries a broad replace patch again, repeating the exact same wasteful cycle on another part of the code or even the exact same source file, doing for 20 minutes what it would have normally finished in 2. It’s gotten a lot dumber. But I guess this also could be a result of aggressive context reduction optimizations. Whatever they did to it, they need to revert it.

43

Seems to be flying faster than ever for me. no degradation at all. Have been pushing it and still not even making a dent in the 5 hour limit. Whatever fix they did I am loving it! Great job team :folded_hands:

44

speed was never the problem its the loss of accuracy and inability to adhere to instructions

I’m reporting that this is still not fixed. It is still making sloppy mistakes that opus 4.8 has to fix

45

After switching back to 5.4 a week ago because 5.5 was no longer reliable enough for my work, I tried 5.5 again today. The issues are sometimes difficult to describe, but I just had a simple and reproducible example: 5.5 generated a text in Markdown, and when I asked it to convert that text into another markup language, it returned the same Markdown text unchanged. I tested the same request with 5.4, and it correctly produced the text in the requested markup language.

This is only one example, but it reflects the broader problem I keep running into. With simple tasks, the issues are already visible; as soon as the task becomes more complex, the output becomes increasingly unreliable and inconsistent.

46

I’ve been following this thread for a while and wanted to add my experience.

I’m not judging GPT-5.5 based on benchmarks or isolated coding tasks. I’m judging it based on a real project that I’ve been building almost daily for the past 6 months.

What worries me isn’t that the model makes mistakes. It always did. The problem is that the nature of those mistakes seems different now.

A few weeks ago I could ask it to modify a specific component, fix a bug, or make a UI adjustment and it would usually stay focused on the requested scope.

Recently I’ve been seeing much more “collateral damage”.

I ask for a small change in one module and it starts modifying unrelated files. Sometimes it refactors things that were never mentioned. Sometimes it breaks existing functionality while fixing something else. Even simple changes now often require multiple attempts before they are done correctly.

The result is that I’m spending more time reviewing, reverting, and repairing changes than I did before.

Maybe it’s just my project. Maybe it’s routing. Maybe it’s context handling. I honestly don’t know.

But my workflow, prompting style, project structure, and expectations have not changed much over the last few weeks.

The behavior has.

Curious if anyone else building an actual production app is seeing the same thing.

47

I fully agree with you. I’m strickly an API guy and use gpt 5.5 for a bunch of use cases. It screams :100:

48

man thats the exact issue i been saying i feel like a crazy person when people keep denying it

i call it ‘splash’ damage’ where it shouldn’t be changing stuff, it is asked to do one simple thing and it touches several stuff, it even reads false positive from tests it writes that aren’t even correct and that kicks off a whole new bag of worms

project hasn’t changed at all but something is going off the rails here. its still happening.

49

After I corrected the result a few times, it didn’t actually make the changes; it just replied, “Confirmed, we’ll proceed with this from now on.”

50

Context comes from a dozen places other than your current thread, so “fairly new convo” is irrelevant.

Each and every skill you have installed, your agents.md , each and every MCP you have install, your memories if you have them, your git commits, etc etc…

Try using GPT 5.5 via the api with zero context and you will see that this issue is caused by your usage and not GPT 5.5 suddenly not working as well.

Its a user issue of course.

51

that’s exactly what I mentioned early on that I started from clear context to measure both and 5.5 did not perform up to par with 4.8

the API and what you use from codex subscription are exactly the same.