GPT-5.5 / Codex quality degraded in the last 2 days, affecting paid renewal decision
Summary
I am a paying Codex user and I rely on Codex for real daily development work.
Over the last two days, Codex / GPT-5.5 has become noticeably worse: more confused, less reliable, and less useful for tasks that it handled better before.
This is not a minor annoyance. My Codex subscription renewal is already due, and this degradation is directly affecting my decision to renew. I am seriously considering whether to keep paying for Codex or switch my coding workflow to Claude Code instead.
What I am seeing
- Codex feels significantly less intelligent than usual.
- It loses the thread more often.
- It needs more corrections for the same type of work.
- It gives lower-quality answers than it did a few days ago.
- It feels unreliable for real daily work.
- The degradation is strong enough that I am questioning whether I should renew my plan.
Why this matters
I use Codex as part of my real work, not just for testing or casual usage.
My USD 200/month ChatGPT Pro / Codex subscription renewal has already expired / is due now, and I need to decide whether it is worth paying for another month.
Right now, Codex feels noticeably degraded, especially with GPT-5.5. Before renewing, I need to know whether this is:
- a known temporary service-side degradation,
- a GPT-5.5 model regression,
- a Codex app / Windows issue,
- an account-specific issue,
- or something wrong with my local setup.
Expected behavior
Codex should remain reliable enough for paid daily development work.
If there is a temporary degradation or known regression, users should be informed clearly, especially when it affects paid renewal decisions.
Actual behavior
During the last two days, Codex / GPT-5.5 has felt much worse than usual and less dependable for the same type of development tasks I normally perform.
Environment
- OS: Microsoft Windows 11 Pro
- OS version: 10.0.26200
- OS build: 26200
- System type: x64-based PC
- Device: HP Pavilion Gaming Laptop 15-dk1xxx
- Codex app version: 26.506.3741.0
- Codex CLI version: 0.130.0-alpha.5
- Codex executable path: C:\Users\Pc\AppData\Local\OpenAI\Codex\bin\codex.exe
- PowerShell version: 7.6.1
- Model: GPT-5.5 / Codex
- Subscription: ChatGPT Pro, USD 200/month, using Codex, renewal currently due
Request
Please clarify whether this is a known issue affecting Codex / GPT-5.5 users, or whether I should investigate something specific to my local Windows installation or account.
This directly affects whether I renew my Codex subscription or move my coding workflow to Claude Code.
6 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Thanks for detecting the possible duplicate.
This looks related to #21375.
I am seeing a very similar regression, but on Windows 11 with the Codex desktop app and Codex CLI 0.130.0-alpha.5.
I have also seen similar recent reports from other users on Reddit, including complaints about Codex / GPT-5.5 becoming worse, slower, less reliable, or failing to follow instructions over the last few days:
The issue is also affecting my USD 200/month ChatGPT Pro renewal decision, which is currently due. I opened this issue separately because I need to know whether this is a known temporary service/model-side regression, or something specific to my Windows app/account/local setup.
If this is considered the same root issue as #21375, please feel free to link or deduplicate it, but I would appreciate clarification because the renewal decision is immediate.
Thanks for the bug report. I'm not aware of any changes that would explain a degradation like the one you're describing here.
If you see more of this, please use /feedback to upload the session details.
I feel the same here
I got similiar experience
---
What version of Codex CLI is running?
v0.130.0
What subscription do you have?
Plus
Which model were you using?
gpt-5.5
What platform is your computer?
WSL2 in Windows10
What terminal emulator and version are you using (if applicable)?
Windows Terminal
What issue are you seeing?
I wrote a tool for my agent, and inside that tool, I left a main function without implementing it, just filling it with pass. Then I said I wanted to see its result, and it actually went ahead and ran that function. What I meant was for it to complete the main function.
Then I saw that the input was written in JSON, so I asked why use JSON instead of pydantic? It interpreted that as meaning the output should also be written using pydantic.
In another conversation, it made error handling extremely complicated, identifying an error code for every single error. I really don't understand what it was thinking.
What steps can reproduce the bug?
Uploaded thread: 019e10da-a691-7ae3-b91a-3b3dd5230136
Same here. Had stable experience with all other models. Can't be coincidence