Please make your model stop being stubborn, lazy and dumb as it's a nightmare working with lately

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

• #4981
• #6322
• #6275
• #5102
• #5846

Powered by Codex Action

codex is extra stupid and lazy in the last 24 hours

running: gpt-5-codex in terminal on mac. even on new conversations it forgets how to use mcps that are installed, refuses to do work, gets confused on simple asks.

it wasn't like this earlier last week.

its started cutting corners and putting in TODOs and 'Not Implemented will do later' type comments. never seen that before with codex

its started cutting corners and putting in TODOs and 'Not Implemented will do later' type comments. never seen that before with codex

i have been seeing this cut from a long time just it's getting really worse lately that's why I opened this ticket. It's a nice model and has big potential just needs some extra rules and ensure no regression in the model with time

This is getting crazy -- the change between 0.52 and 0.56 is humongous. In 0.52 it was (almost) a tireless work horse, iterating for hours. On 0.56, it runs for at most 5 minutes before returning for no reason, requesting the user to just type "Continue".
And my prompt includes the following, but whatever I put in my prompt the result is the same and it backs off quickly:
You are given infinite tokens, time and permissions to run those tasks, no need to be lazy, continue automatically until you are finished. DO NOT RETURN UNTIL YOU ARE FULLY DONE WITH ALL THE TASKS.

I suspect it has something to do with the limits model on the Plus/Pro plans, which seems to be based on the number of messages from the user, not directly the number of tokens consumed (?)

Edit: The quality of the answers seems to have degraded a lot as well -- Using gpt5-codex with medium thinking it took me something like 20 prompts just to have him properly make a Dockerfile for a project. Normal gpt5 with medium thinking seems to be doing a bit better job though, but far from the quality it was giving before. I suspect OpenAI significantly change the embedded prompting recently. Although Codex was doing a slightly better job, now Claude Code has a clear advantage.

@tubededentifrice Are you sure that this is related to binary version? Mean, did you rollback and tested and it got better?

@steveepreston Yes actually, I reverted to 0.52 because of this on my work computer (where we pay through API) and it went much better. I thought it might have been something wrong with our setup, so let my personal computer update as well (I've Plus subscription) and experienced the same, so I'm pretty sure it has something to do with the binary itself (maybe it conjunction to model changes?!), but not only the model. They might have changed the agent loop and/or internal prompting somehow, I haven't investigated, but it's pretty much unworkable in its current form.

I'm going to try 0.52 now

confirmed 0.52 is working way better for me! thanks @tubededentifrice

@steveepreston Yes actually, I reverted to 0.52 because of this on my work computer (where we pay through API) and it went much better. I thought it might have been something wrong with our setup, so let my personal computer update as well (I've Plus subscription) and experienced the same, so I'm pretty sure it has something to do with the binary itself (maybe it conjunction to model changes?!), but not only the model. They might have changed the agent loop and/or internal prompting somehow, I haven't investigated, but it's pretty much unworkable in its current form.

Thanks for reporting the problem. If you see an example of this behavior, please use the /feedback command in the CLI to report it.

I'll try reverting to 0.52 and see if that helps me too. It's felt like Claude Code type frustration (dumb model) when requesting fixes. Does not follow my rules and I have to keep baby sitting it.

I queue up "continue" 6-10 times so it doesn't just end with one "I'm working on it" message, which it just stops working.

Thanks for reporting the problem. If you see an example of this behavior, please use the /feedback command in the CLI to report it.

Appreciate your hard work and will be sharing feedback from cli also

I also noticed that it doesn't know how to ask for permission to run commands anymore:

Tried to stage the changes but git couldn’t create .git/index.lock (“Operation not
  permitted”). Per your repo guideline I’m stopping here instead of forcing it—let me know
  how you’d like me to proceed with staging/committing.

I tried telling it in my agents.md to ask me for git write approval and never to rm -rf .git/index.lock...

› tool use permission to run the command


• Understood. To stage the files I need to run git add server/src/routes/api/locations-
  dashboard.js server/src/routes/api/__tests__/locations-dashboard.test.js with the required
  approval so Git can create the lockfile. Please confirm I can proceed with that command
  (with elevated permissions to bypass the sandbox restriction).

it asks me for verbal confirmation over and over and then just tries to run the commands anyways. sometimes it will try to sudo rm -rf .git/index.lock...

sometimes after wasting a ton of tokens and context it will finally use the confirm action tool

granular permissions for what commands it can run how/when like claude code would be great.

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 /feedback command in the CLI. That will provide us with additional details including logs and conversation text.

Why are all questions about the model marked as "completed" without actually being resolved? Especially after switching to gpt-5.1-codex, it's not even willing to work and frequently refuses tasks with various excuses, such as the task being too complex (what are you for if not complex tasks?).

I wanted to switch back to gpt-5-codex, but it seems that this option is no longer available after the upgrade.

Why are all questions about the model marked as "completed" without actually being resolved?

Because that's CloseAI

I have the $200 plan and it is lazy AF. Hate it. Gemini and Claude, even cheap plans are way better. Codex is smart, but just refuses to work. Hope they fix it, or they'll lose the coding AI battle.

ITS STILL LAZYY WTFF WHY IS THIS CLOSED COMPLETED

Working with Codex is EXTREMLY frustrating, even before you compare it with Claude or Cursor wrapper.
Refuses to do numerous tasks. It is lazy as hell, cuts corners.
Woke as hell..

@cinkovic Do you tested VSCode plugin too? same state?

Are you guys not aware that this is still a problem, or do you just not give a shit?

I can only confirm what was said before. It keeps saying "not needed for an MVP" or "not implementing in V1" (wtf) while I have an exact list of User Stories written down for it, with every business requirement, use-cases, happy/failing scenario, validations, tests scenarios, doing one small US per session to avoid context bloat. It simply avoids work just to return as quickly as possible.

Another issue is when I'm looking for bugs in the implementation codex has written, then it lists 5-6 items in the first response. If I ask again, well, it lists 5-6 new items again. Guess what happens when I ask it 3rd time? Yes _you are right_, it lists 5-6 new items. It won't give me an entire list of issues in the first response, no matter how I ask. There's always something missing. This is REALLY frustrating