significant reduction in intelligence. feels like gpt 5.5 got downgraded to 5.3.

Open 💬 7 comments Opened Jun 25, 2026 by sorcrr-dev

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.623.30605 • Released Jun 25, 2026

What subscription do you have?

pro $200

What platform is your computer?

_No response_

What issue are you seeing?

the last two days gpt 5.5 doesnt feel like 5.5. the quality of output, the design results, the context/memory about the product have all been significantly degraded to the point where if this continues im probably going to be unsubscribing and moving away from openai.

What steps can reproduce the bug?

created a new screen on flutter with reference to a screen created 2 days prior asking the model to use the same pattern. instead of getting the same output quality got something 5.3 would output; opus 4.1 level.

What is the expected behavior?

expected behavior was to leverage the memory and agent.md files like it was before, be able to use defined patterns used before to adapt new features to new screens. not having to go back and fix the issues and iterate 10 times on something 5.5 was doing in what shot before.

Additional information

started 2 days ago

View original on GitHub ↗

7 Comments

github-actions[bot] contributor · 25 days ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#29879

Powered by Codex Action

QuinnISHE · 25 days ago

I've also had this same issue, 5.5 works much slower, doesnt follow instructions, even has begun intentionally weakening tests before running to get them to pass, and takes much longer to actually complete my specifications

UEhQZXI · 24 days ago

You can see which model you used in Codex from the Codex Cloud Analytics page, under By model:

[https://chatgpt.com/codex/cloud/settings/analytics](https://chatgpt.com/codex/cloud/settings/analytics#:~:text=2%2C968-,By%20model,-By%20surface)

sorcrr-dev · 24 days ago

i checked out the analytics. It shows mostly 5.5 but we all know there are many ways around that. The issue is more along the lines of performance gpt 5.5 could do anything before - it literally felt like coding was cooked - thats how its been since I switched from anthropic to openai. Superior problem solving ability, would find issues that I wasnt aware of and fix those along the way. All of a sudden its just flipped back to 5.3 times. I have been stuck in the same position for 3 days now, with the model not able to implement something that it would take an hour to do prior to this drop in quality. The output is significantly poorer. The first shot of anything is like how things were back in 5.3 codex days. The model makes stupid mistakes, much higher error rates, much worse at following instructions, weaker memory/context recollection. It cant even copy established design patterns, after you have it audit it, read the design markups. What has changed is that now it runs for longer but if its output is weak thats just more time wasted. Today is the third day of this. Its also a lot slower. Will switch over to Opus 4.8 and see the difference. Maybe give glm 5.2 and kimi 2.7 a try as well.

obcardinal · 24 days ago

Yes, exactly — it feels like version 5.3 instead of 5.5.

To be fair, I only started using Codex from version 5.4, so I cannot directly compare it with the real 5.3 experience. But what I can say is that the current quality does not even feel like 5.4 anymore, let alone 5.5.
However, I also do not rule out the possibility that, even though I always use reasoning at xhigh, it is actually being executed at a low level.

But something has definitely changed over the last 2 days. 5.5 (with xhigh) misses a lot of details, makes basic mistakes, and does not verify the result. Even when the task is described in detail, it can still simply skip important instructions.

I hope OpenAI has not started limiting the number of tokens consumed per task, causing the AI to take the path of least resistance / the easiest route (aka Anthropic with Opus 4.6).

Right now, I even have to handle ordinary work tasks under a parent 5.5 Pro, which supervises Codex 5.5 and suggests further prompts for solving the task. I cannot say that this helps much at all.
When the problem is clearly located in a single file, giving it to 5.5 Pro allowed it to immediately find the narrow bug in the code. Codex 5.5, on the other hand, started rewriting half of the file…

pandigita · 17 days ago

26.623.81905 / 5.5 Medium

Same here, it's making all sorts of bad decisions. Quality of output is drastically reduced. First time I've experienced Codex having such a massive downgrade, though no stranger to it in chatgpt.

I'd also like to point out that there is a real psychological impact to having your thinking partner lobotomized like this - even when you only use it for work. It's seriously impacted me before in chatgpt and led to drastically reduced usage on my end.
When you chose to lobotomize your models for cost savings or freeing up GPUs, please don't just think about the impact on your users as "intelligence utils lost" - it's much more than that.

YanhaoZhang · 13 days ago

I asked the last date the model was trained, and it turns out Jun-2024. This may explain why we feel the model is much more dumper than before. But it is so unfair that OpenAI downgrades the model without any notice. Besides, there is no announcement when the new model is available. Besides, I am a Pro member. I do not know, it just happened for Pro members or it happens for all of the memebers.