Codex usage page shows unstable remaining 5-hour quota

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

• #32029
• #31696
• #32129
• #31758

Powered by Codex Action

I can report, I noticed something simular --- the 5-hours need to be sent twice (two times need to type /usage or /status ) now its instant

I could report the issues with the indication ("How much usage I have left") is fixed ... per/by how much per 7d limit - it vary?

as per "fluctuate unexpectedly" - I did noticed it too -- there is evidence they(Open-A.I) are expirementing with the ratelimits -- nothing seems weird now -- despite I can report the speed and ratelimits did catched-my-eye

present time @Jason-Wang313 Does this feel as something fixed? for me I consider this as "fixed" and that openai dealt with the areas of this issue. a month after you filled this issue -- this section of codex /status (or quota-limits/allowance)
@Jason-Wang313 I think the issue is fixed - i felt "unstable remaining 5-hour quota" but in many ways what I noticed isnt there now - so for me ts fixed -- do you have see the same thing?

I was confused too sometimes I drained all of the 7d limits, and then it gets a reset (a reset means, they reset the 7d limit for all Plus or Pro users) (if they can afford it , why doesnt it become a 5d limit? Should the user treat it like gold - and be very careful with the allowance? so more days and more risk you waste your badget== the better?)
Now I am not confused -- for me with the launch of 5.6 models - they are some growing pains - but now its more matured? i dont feel any issue with it ---

for me I type /status or /usage -- and nothing execpted happens -- which wasnt before - etc need to retype the commend 2 times /status --
for me it drove me to reach all of my limits after reach reset - because of the fear of missing out - if you work on many project I recomend try to optimize yourworkflow to only use gpt 5.6 luna , if a task fail - try to learn on the topic - and then you use ai? (try use a search Engine like Yandex or Google or Bing --- or Chatgpt to find sources - trying to find "Keywords" related to my topic, and learnn) --> on then use your coding agent? but you should surely learn on the topics you tasks to do --- Whether its a task for some purpose --- or a task like that a user face or use (or a user is a effected by) ... how often is it going to be used, or is it to the trashbin (unless you fix something - or optimize)... prior knowledge, means a more concise and fast execution of the goal. despite other factors calculated here --- but for example compilers have bugs, thats why we trust rust - but honestly there is so much to do nowadays even with what we consider little - instead of getting confused of noise and blaming others -- for examples cpu-s are faster than ever compilers are better than ever, for example intel and amd over the compilers - even in areas of the compilers that barely change (but for examples Debuggers are a very big deal because compilers have bugs) - instead of extended features as time progress - I have big respect to javascript or typescript - but I like the least amount of lines, and the most amount of dependecies unless I get tasked to learn a small technicallity

Maybe try watch another friend /collega/worker/employee who use it --- slowly watch maybe you can adopt a better way - in the use of the tool..... We all get the same vibe - from vibe coded websites? or software? so the use of AI to build something is lacking - and grounded - or lacks personality? --- as per efficency - it comes to prepare with what you want to change or fix --- what you want to remove --- or actually must remove without your wishes (yes). but clearly sometimes you need to deal with a lot of files -- the and user wish not to change the bad way of having too many files (legitimate or other reasons such as legacy or priorities) ----
programming is the propurse of serving an issue or a user? or just technical maintaince? sometimes a wrapper or an extention to something. sometimes a task to learn the mechanics of something ---- maybe you should move a tier, use pro and not "plus" what deter you. does it drain faster than your expectation so it deters you?
as per why prompts can fail? why "use cases" fail , interesting issue. for the most part I personally get things done with it - but some tasks fail - until they dont fail (took the user to make few prompts -- "repaired-threads" users added prompts/messages to something that failed --- what are the clues - the user used "undo" inside the ide over 10k lines commit --> "I am about to make 100k lines letters change, are you sure you want to do this?" ) in the future this should be reduced(the amount of turns/failed-threads) -- now we have/had seen more use cases --- we adopted.... on another node you see some wasted usage like inside the normal chatgpt version - you ask it to go to a source /or/ crawl something - and then in the tool usage in the cloud version of chatgpt - you get something like "Premission denied" because it didnt used sudo or didnt have library or something like that? but it still does what ever wanted it adopted the next turn still got we wanted --- but dozen of users have kind of issue ---- and I dont think its perfect -- if we didnt saved 20second , we stalled a gpu - and the same user wont make another chatgpt chat for the next 2 hours - therefore we should been quicker with that user/client - due to the fact we might stalled a gpu/resources ? (this is something when I used gpt 5.6 sol on the chatgpt version / chat / [not work/codex] ) --- on another node, you can run the same prompts 2-4 times, and sometimes the out put is a bit different/completly different - which is magic for me - as things stands now there is no "prompt enginering" it seen all of the bad prompts (personal opionion) there are just slower prompts--- in this point of time (i say, basically "there are faster promtps") --- writing to it in none-english does seems to be a bit faulty -despite I think gpt5.6 I heard was an improvement - sysetmprompt in english but you will communicate with me in Chinese or Korean, but I cant commend how bad it is or not - I personally use english to prompt to get the most juice, If it takes more tokens if you use chinese or korean or the same I dont know, I also dont know the count of turns/threads/comments/threads it makes, if the language itself differs ---- or how present the intent of the user is - based on which language he chose to speak - how do you event take the threads, and see which threads are a success or not? the answer is technical but should be answer with basic human behavior - and the goals you wish to achieve and a combo of best/likable --- speaking of the best of the best -- for now there is a lot of turns -- the best thing to do would maybe use or download a thirdparty tool - or utalize (likely an already ecixting command inside the unix/gnu utils -- for windows users Powershell [since few months ago this actually improved as I was very happy with 5.3 or 5.2 or 5.4 ] ) --- but the model will rewire a small python/nodejs snippet - due to the training process --- or some human behavior users will say no I dont want to use a tool debicated to the niche usage - so you should apply to what the user want --- you can also argue- the model should argue "are you sure, lets avoid this, use this tool - this will be quicker for us - [explain in a word]" -- and again, as things stood few months ago even, I cant be more happy than we got in previous iterestion