Codex usage page shows unstable remaining 5-hour quota
Resolved 💬 3 comments Opened Jul 11, 2026 by Jason-Wang313 Closed Aug 6, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
Environment
- Windows
- Latest ChatGPT desktop app / Codex integration
Problem
On the Codex usage page, the remaining quota for the rolling 5-hour window appears to jump around unexpectedly.
For example, the remaining value may change from 100 to 67 and then back again without any obvious user action.
Expected behavior
The usage indicator should be stable and clearly explain how the remaining quota is calculated.
If the quota is recomputed because of a rolling time window, concurrent tasks, or delayed accounting, the UI should indicate that instead of making the remaining count appear to fluctuate randomly.
Why this matters
The current behavior is confusing because it is difficult to know how much quota is actually available or whether usage has been charged correctly.
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Having worse problem. Mine Codex reduces 5 hour limits by its own even without any model work done. Just started the app and browsed through some local chats and got wasted around of 5% of 5 hour limit =(. Have feeling that there is some bug and current app consumes limits by time not by the work done.
thats the price of having 4x the usage as 5.4 launch, which as a bigger model than 5.3, and meant to be expiremental considering how few people used the codex itself - now its a market leader, I guess I dont want to see all of claude users migrated to codex because this would mean having use more hardware that doesnt exist, I praise codex so much because they ave smaller model , whats the point of gpt sol 56 if it cant serve so many users --- on the other hand gpt 5.6 luna, is the same strength as 5.5xhigh.. and lets be honest all of you people like 5,5 because you can get things done with 5.5 ....
instead of prasing the stength of older models, (and a lot of what 5.5 had model-behavior issues and what it seen, now 5.6 learned from what 5.5 has stragled with ) instead of trying to compact the power of 5.5 , we launch 105 agents , which could have mean 20 users using the same capicity
but that being said, I do consider the PLUS subscription I use, literally unlimited - but I dont like using SOL . no point of using it I guess I liked using /review over it
I think whatever magic is behind the 5.5 models , is that it was powerful but the data was lacking --- now it seens more use cases - and it knows to deal with them faster
(as per anthropic , they ratelimit users if they are in china -- something openai never has done)
(as per anthropic, they cant get a smaller more efficent model, the must make the model talk for long before doing anything.)
I just dont see the point of even using SOL, when we got Terra and Luna.... but evals dont lie they do score higher --- but I think certain tasks break the evals --- for example its not great on biology partly because its dataset itself is highly debated and a debate -- everything in bioloy is going to be a theory (a group of facts perhaps, but can vary) , (should be give Amphetamine to kids, Or just Rithalin --- as it just what you happen to prefer) as with llms they are meant to be general for a lot of tasks, I guess it can be overfit on some tasks, but it must be logical when it comes to tasks and data it never seen