This topic is for discussing Codex limits after the move from message-based usage to token-based usage.
Additionally the 2x rate limits promotion has ended and some users have reported hitting limits sooner after this change. This is the right place to compare experiences, share what seems to trigger it, and discuss what helps make usage last longer.
Under the new system, bigger prompts, larger repos, long context, large AGENTS.md files, MCP servers, Fast mode, image generation, and model choice can all use up limits faster.
This is why people hate tech giants. The pattern is always the same: hook users, make them build around your tools, then start squeezing usage, limiting access, and calling it progress.
From the outside, it looks greedy.
From the user side, it feels worse: unstable, manipulative, and arrogant.
You do this, then wonder why competitors keep gaining ground.
i’m on a plus plan. i managed to wipe out my weekly usage in a few hours using 5.4 medium in intellij idea. i’m a little lost on how this can possibly be worth paying $20 a month for? the work it accomplishes is great, but it lasts very briefly.
I’m almost at the point of cancelling when the rates can burn through that tasks that were fine to do a few days back etc are now chewing not only 5hr rate but weekly rate .
Just been reading through all the messages on his topic. I was lucky, I had just 2 days of really burning through my allowance much faster than usual. Since then it appears to have settled down to be more like it used to be. However my weekly is quite alarming and does not reset for another 4 days. I mainly wanted to mention Fast mode. I use the Mac App for all my codex work and a few days ago there was a pop up about enabling fast mode. I was probably naive and did not read about Fast mode properly but I just switched it off in settings when I read that it uses 2 x more usage. I can live with responses that are slower if it gives me better usage. It gives me time to think as well! I absolutely love Codex and do not want to be without it.
No subagents, no cloud work. Im hitting the 5hr window limit within 20 minutes into a light codebase diagnosis + refactor. No major refactors, no exhaustive codebase exploring. Its unusable, hitting the limits not even halfway into a plan.
There are many of us, and even 100x more of those who don’t post here.
Hope OpenAI will fix this next week…? It’s an amazing tool, but unusable at all in current scenario.
Why is my GPT Pro subscription hitting the Codex usage limit so quickly? This never happened before this week. In the past, I didn’t even reach 50% of my weekly usage, but now the usage is completely inconsistent with what I would expect from the Pro plan.
Just wanted to chime in to state the reality of this issue. I have a team of 5 and codex is unusable with these limits. These are not hard large tasks we use codex for, the usage limit drain from quite simple tasks is palpable.
We will switch provider if this is not solved soon. All the best everyone.
I activated six business seats a few weeks ago, paying upfront for a full year. Over the past week, it has become much less usable. After just a few prompts, the 5-hour limit is reached, and the weekly limit runs out halfway through the week, without any particularly heavy usage or complex prompts—just ordinary use. With the same settings and the same model, just a few weeks ago these limits were only reached after very intensive use. I hope this is just a temporary issue; otherwise, I’ve truly wasted money on six annual business subscriptions…
Hello! I also faced this problem. I had paid several business accounts, and the limits began to be spent too. For example, one simple request consumed 75% of the 5 h. After reading this topic, I realised that it was the users of business subscriptions. While the owners of plan plus or pro do not face this. It’s very interesting when this problem will be solved.
Actually, I have tested this myself, and business usage is depleting much faster. A large part of that does appear to be tied to the recent “known” changes. That said, Plus personal accounts are still being affected as well, just not to the same extent.
What is interesting is that, at this point, you are arguably better off with a Go account and its weekly quota, simply because it still allows normal, practical use.
I have also been doing some digging, and it seems quite possible there is a separate issue with Codex and/or the IDE extension. If my testing is correct, a few very simple requests were consuming around 16M input tokens and producing roughly 100k output tokens. That is nowhere near normal, and it would certainly explain why token balances are dropping so quickly.
I think I’ve figured out the reason why the limits are being consumed so quickly. However, I’m still not completely sure about its accuracy.
Every time Codex does something—reads files, searches through code, runs commands—it receives output and adds it to the history. The history grows with every action the agent takes. And every subsequent request you send, even if it’s just three words, is sent to the server together with the entire accumulated history.
So if there were three large file searches during a session, that’s already around 600k tokens sitting in the tail. You type “continue” — you send 600k. You type “okay” — another 600k. And so on, until you close the thread.
In other words, it seems like the issue might be with how many tokens are being sent with each request. That’s what’s causing such high token usage and, consequently, hitting the limits so quickly. However, this is just a hypothesis, because I’ve noticed some anomalous behavior. Even in a newly created thread, each new request sends an unusually large number of tokens that shouldn’t be there. And most interestingly, the number of tokens being sent keeps increasing with every request.
I’m on Pro. Until about mid march, I could work on coding projects for hours every day, with the occasional 5h cap. I usually never ran out of my weekly. One day it just changed, and I used up 100% in one day without any noticeable difference in my workflow. I don’t think this is intentional, I think there’s a problem or a mixture of problems at play. The really weird thing is blowing through my entire week before I hit my 5h cap, which happens.
Miscalculation of consumed token consumption?
Caching broken or how the cache is calculated related to token consumption?
Context multiplication, unintentional bloating of context?
Or maybe the calculation was off BEFORE and they fixed it and now reality is that inference is really expensive and machine intelligence costs more than we were used to.
I’m not here to complain but I am here to say that something changed by an order of magnitude recently such that I can’t rely on the tool any more. I try to never use 5.4 anymore as ever turn can blow up like 1-10% of weekly limits.
This also makes sense as from what ive been observing like simple commits are using up excessive tokens. Even very small refactors like removing X files from a dir is causing massive token usage and its even worse if its looking in files for stuff to remove.