Repository navigation
Rapidly hitting weekly limits on ChatGPT Pro plan #3734
Description
What version of Codex is running?
codex-cli 0.36.0
Which model were you using?
gpt-5-codex high
What platform is your computer?
Linux 6.14.0-29-generic x86_64 x86_64
What steps can reproduce the bug?
Use codex as your primary coding agent.
What is the expected behavior?
Usage parity with Claude code. Codex usage limits are substantially lower than CC.
What do you see instead?
Incentive to not spend on ChatGPT Pro
Additional information
it is caching a tremendous amount of tokens per chat
Token usage: total=748,177 input=693,709 (+ 9,781,504 cached) output=54,468 (reasoning 33,216)
Maybe this is why?
No response
Activity
ariawisp commentedon Sep 16, 2025
gpt-5-codex high seems to consume insane amounts of usage compared to gpt-5 high. I had been using gpt-5 high with 4-6 agents for 8+ hours a day on my Pro plan without ever hitting usage limits - then I tried gpt-5-codex high when it released yesterday and hit the 5-day usage limit after only 30 minutes of running 2 agents. Very frustrating (and agreed: What do you see instead? = Incentive to not spend on ChatGPT Pro)
aehlke commentedon Sep 16, 2025
Since it released they heavily recommend using medium. It's more dynamic than before - using fewer tokens for thought when it identifies "easy" problems (without using a router), and using more tokens (like high) for "hard" ones. You might want to try this.
Btw how many tokens did it report were used in those 30 minutes?
ariawisp commentedon Sep 16, 2025
If I recall, it was around the same magnitude shown by @Alfredo-Sandoval - around 10 million cached tokens (far higher than I ever saw with GPT-5 high)
aehlke commentedon Sep 16, 2025
Sounds like something is going wrong then. I used around 10 million un-cached tokens yesterday, on my first day of Pro usage. This was using medium though.
From my largest session:
Token usage: total=8,755,957 input=8,537,903 (+ 30,605,824 cached) output=218,054 (reasoning 122,496)Alfredo-Sandoval commentedon Sep 16, 2025
Had the weekly limit lifted for a few messages and its back now. Can't help but think its related to the usage spike from people coming over to codex
TrevorLoucks commentedon Sep 17, 2025
I just hit limit on Pro plan using gpt-5-codex-medium with 2.5 days till reset, something is off. Max 2 terminals at a time this week so far using only medium.
And the limit message is also telling me to upgrade to Pro to increase my limits.
Latest session stats:
Session ID: e1a372f3-fc9e-49fa-af2e-6b894b7d5bce
• Input: 0 (+ 12929408 cached)
• Output: 44,736
• Total: 44,736
ariawisp commentedon Sep 17, 2025
Tweets:
This doesn't seem normal. Hopefully OpenAI will address this ASAP - 2+ day timeouts is unacceptable for those of us paying $200/mo and relying on Codex as a crucial part of our workflows.
omggga commentedon Sep 17, 2025
Unfortunately, the docs only mention a weekly usage limit but don’t specify the actual numbers. Since the gpt-5-codex release, it feels like the cap has dropped significantly. Could you please publish the concrete weekly limits (per plan/model) and reset rules?
Hitting the limit after half a day of normal work on a Pro plan is pretty unreasonable without clear guidance.
rockyicer commentedon Sep 17, 2025
you scare the shit out of me. I am using high sometimes , most time is medium. now I will watch my back when I use high. thx to mention this.
ariawisp commentedon Sep 17, 2025
@rockyicer I've just gone back to gpt-5 high and having no issues with limits there so far, avoiding the new gpt-5-codex models entirely until these usage limits issues are confirmed to be fixed.
gpt-5 high is working better for me anyway, in the limited time I tried gpt-5-codex the other day, it spent 30 minutes working and ultimately reverted everything it did because it wasn't sure how to solve the issues that were coming up - never saw that behavior with gpt-5.
iwangjie commentedon Sep 17, 2025
Does it seem too smart?
ariawisp commentedon Sep 17, 2025
@iwangjie I don't know if I would call it "too smart" but from what I've observed it seems aggressively RL'd in such a way that it views itself as more of a solo engineer / project owner than an assistant/collaborator.
whoschek commentedon Sep 18, 2025
@iwangjie In my experiece gpt-5-codex high also turns out to be far worse than gpt-5 high, at least for a complex codebase. It comes across as a shallow drama queen that claims to find bugs where there actually aren't any. The bugs it sees are simply misunderstandings on the part of gpt-5-codex high - it's thinking too shallow. gpt-5 high doesn't do that - it performs very well and it does so consistently, and I'm using it a lot (and before that codex-1).
My speculation is that gpt-5-codex high is actually a very capable model (and probably better than gpt-5 high in a theoretical sense, or in some specific cases) but the part that makes it auto-decide how short or long to think is often seriously misbehaving. So the model ends up spending little reasoning time on things that are actually subtle and hard issues. I find that this model behavior is definitely reproducible, it's not just a one-off, or rare. Inconsistency kills the UX.
I'm more and more convinced that the real fix is for OpenAI to drop the "auto-mode", and instead add a radio group UI widget where users can choose between think time: 1sec, 10secs, 100secs, 1000secs (or even a log scale slider that allows all values between), because the user knows much better what he wants/needs than the model ever will.
jovkon123 commentedon Sep 18, 2025
Just tried first run of gpt-5-codex medium and it spent 45minutes running powershell cmds changing 2 files and then ultimately reverthing them at the end. Incredible waste of time it returned this message:
codex: I’m sorry, but I wasn’t able to finish the implementation.
📊 Token Usage
• Session ID: 5f0463d4-4eb2-4527-9348-6436342a8cce
• Input: 267,626 (+ 17749504 cached)
• Output: 117,325
• Total: 384,951
11 remaining items
aehlke commentedon Sep 24, 2025
That's troubling - I haven't run into it yet
Does anyone know how the weekly limit rollover works? I can now see I'm at 38% of my weekly limit, but I have no idea what that means in terms of my pace. E.g. am I 38% done just 2 days into the week, or have I underused it and it's going to reset tomorrow? Or is it a sliding window which would mean I should try to stay near 100% without hitting it to avoid wasting unused tokens? I haven't been able to find any info.
edit: #4080 covers my question (unanswered though)
spartan-it commentedon Sep 24, 2025
Well guys, Let me show you something even crazier. After my limit had reseted I tried a single prompt in gpt 5 codex medium and after waiting for 30 minutes, the AI stopped halfway.
babyghost-ys commentedon Sep 28, 2025
I didn't use the high, only using the default medium. Reaching 75% within 3 days. Very disappointing. People were complaining Claude Code Pro US$20 also has the same issue, but I am 99% sure it's not. I was working only on one side project with Claude Code last month, very similar usage per day. Never reach the limit of Claude code.
Then, this month, switched to Codex and this quickly happens within 3 days. Cancelling the subscription for now until Codex becomes more reasonable.
tekn0x commentedon Sep 28, 2025
I used 10% of my weekly Pro limit coding for 8 hours with "gpt-5-codex high" with a single agent in Codex CLI. I closed and reopened a bunch of times, but just 8 hours of feedback loops. No crazy agentic mega task lists or automated agentic workflows like Speckit or Bmad. Just normal manual back and forth.
So $200 a month is cutting it very close at that rate for a 5-6 day work week. At least claude code never had the weekly cut off, just the 5 hour cool down. Any one using multi agents or GitHub actions would be cutting it even closer.
neo0oen619 commentedon Oct 5, 2025
like really i just had 5 hours of coding and i get limited for 5 days .... that absurd
"■ You've hit your usage limit. Upgrade to Pro (https://openai.com/chatgpt/pricing) or try again in 5 days 19 hours 29
minutes."
at least the new version is telling us 75 90 and 95% use ... really im just getting of sick trying to do stuff like this .. :[ sam Altman and codex please read this and fix
aehlke commentedon Oct 5, 2025
You'll need to either sign up for multiple Plus accounts or upgrade to Pro. They're unlikely to raise it much from here since Pro is pretty good for nonstop usage without parallelization now
fanxxai commentedon Oct 8, 2025
The usage limit must not be this small.
And when the limit is hit, an alternative model should be available.
ariawisp commentedon Oct 14, 2025
This issue has been open for nearly a full month and still no acknowledgement from OpenAI. I've been using (well, attempting to be using) Codex on my Pro plan exactly as they advertise it:
Yet, I'm still reaching the weekly limits multiple days before the weekly reset takes place - a direct contradiction of the experience they state Pro users will receive at https://help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan
How can I "power [my] full workday across multiple projects" if attempting to use the plan exactly as they state it can be used results in me being locked out for multiple work days at a time?
javothemes commentedon Oct 24, 2025
I mostly use "medium" in Pro. But it goes really fast. I am a mid-heavy user ( not very heavy user ). but always hit near 100% in weekly limits. Never hit 5h limit even 50%. Not many mcp ( only context 7. not much used )
I believe weekly limits should be more limit.
etraut-openai commentedon Nov 11, 2025
We're recently introduced some changes that help your extend your available usage.
mateususa commentedon Feb 1, 2026
Introducing a less capable model does not fix the issue. If OpenAI had a tiny model that was capable of auto selecting an appropriate LLm given the task's complexity that would be great for everyone.
etraut-openai commentedon Feb 1, 2026
@mateususa, our most recent models already dynamically adjust their reasoning level based on the specified task. We're continually looking for ways to lower the costs of inference while increasing what you can accomplish within a given budget. If history is any guide, you can assume that costs will continue to drop over time while capabilities will continue to increase. If you are hitting usage limits in the meantime, you have the option of upgrading to a less restrictive plan, buying additional credits, or switching to metered (pay-for-what-you-use) API key pricing.
mateususa commentedon Feb 1, 2026
Sounds great. I was aware that supposedly the Gpt 5 models are meant to auto adjust. But from my experience. At least in the Web app (I am new to using CLI) I have not experienced this auto adjust. I always keep "extended thinking" ON because the regular 5.2 never seems to satisfy my questions. I suppose this happens because there is not a lot of good quality data for the model to be trained on. Thus I seem to always need extended thinking ON for the model to take edge cases into consideration.
Anyhow, it's hard to complain about a great tool.