Change rate limit accounting
Resolved 💬 18 comments Opened Sep 25, 2025 by 239573049 Closed Jan 31, 2026
What feature would you like to see?
Why are the weekly limit and the 5-hour limit so close to each other? I would like to ask you how you plan to compete with Claude Code? I don't think it's necessary for me to subscribe.
<img width="644" height="123" alt="Image" src="https://github.com/user-attachments/assets/5b24bb4b-3e73-40f4-b5ac-d94c4e0022db" />
Are you interested in implementing this feature?
_No response_
Additional information
_No response_
18 Comments
That is to say, if we exceed the 5-hour limit twice, I won't be able to use Codex this week.
Here’s the translation:
On Windows, MCP support is really poor.
You hit the weekly limit after only a few uses.
The CLI optimization is terribly bad.
Especially with the Team version—you hit the red line (quota cap) after just a few tries. This is, in my view, the most ridiculous subscription I’ve ever purchased!
“At a price of several hundred dollars, yet I can use it for less than a day each week—are you robbing me? And how are you even supposed to compete with CLAUDE?”
They are close because the weekly limit is VERY VERY LOW.
There is no point to using Codex. It is basically a tech demo to keep them in conversation since Claude Code is all that was being talked about.
Claude Code has way more value and much better in every way.
For my tests
5h limit x 3 ≈ 1 weekly limitits very lowI described what it probably stems from.
They are not close to each other even if they would show the same number lets say 34% , one is 34% of weekly (bigger amount) usage, other is 34% of 5 hrs (less amount of something) . Results are still not same , because totals of those 5hrs and weekly usage limits are not same.
Just 2 independent metrics there. no proportion between them. You can see 5 hrs limit is 99% and weekly is still 39% .. Or do you mean something different ?
The point is that the ratio is stated at the top and clearly described as 1:3, not 1:1.
No one claimed it was 1:1 — it’s explicitly explained why the ratio is low and not equal.
So, what you’re saying is both true and not true at the same time.
It depends on the ratio, and as you can see, that ratio is low.
And this is caused not by the limit itself, but by token burn.
With the same limit, proper context management could increase the number of requests by 10x–50x without touching the limit thresholds.
My guess is that Claude has lower raw token limits, but better context management, which is why it achieves that effect.
So, to answer the original poster’s question: if they want to compete with Claude Code and others, they must improve their agent.
They simply can’t afford to burn through limits so aggressively — doing so would cost hundreds of billions of dollars in GPUs.
Send single messages and check /status after each one — you’ll see the cache size grow quadratically.
Originally, the problem was masked 7–9x(and that’s why the limits increased by that much), because they were recomputing the entire context at every iteration.
Now, the input only shows you the new tokens, but that just hides the issue, because cached tokens still cost money.
And at the scale of such an explosive growth, having cache 10x cheaper doesn’t really help much.
Here is the thing. I purchase additional credits, that is ok and the cli version on wsl2 was using the credits. Now all of the sudden a hard limit that cannot proceed with the work i am doing even if i pay for it because of this weekly limit. The best part is that the cli says that the weekly limit will reset in 5 days. What you guys suggest i tell to my client that i have to wait 5 days because somebody decided for god knows what reason that i cannot even use the credits i paid for? I am using the tool for real work. Thank god there is competition. Last but no least i ask in the "support" chat and the AI (again) tells me that a human is going to take "days" to answer my query. So welcome to the AI blues. Hope that an OpenAI human is reading this. If i cant use my credits for the codex because that is what i buy them for I Want my money back so i can go to another provider and get the service. the best one is that the system is actually using my credits...
I’m a Pro user. Right now, with a 5-hour limit, my weekly limit increases by 1% for every 2% used. If that’s the case, if I use up all the 5-hour sessions in a week, does that mean I can only use two sessions? This seems very different from the statement that you shouldn’t hit limits unless you’re doing scripting. I don’t do any scripting at all; I just keep one or two windows open and work slowly.
that's crazy...
Just now I got suspended by weekly usage 100%，but my 5 hours usage are still 19%，I cant get the point... how did it happen and what's their calculation way?
Why is there even a 5h limit?
Why can't this be removed?
Also the weekly limit.
It makes no sense other than annoying users.
If we have a monthly contingent of tokens let us use all the tokens and be done with it for the month, or let up buy more.
But this 5h and weekly limit is BS.
https://blog.langchain.com/deep-agents/
That’s exactly what I was thinking and talking about.
[https://blog.langchain.com/introducing-deepagents-cli/](https://blog.langchain.com/introducing-deepagents-cli/
)
In two days ive managed to hit 77% of my weekly allotment. Leaving me out of luck for 5 days if i continue. Im a plus subsciber. Claude has no weekly limit. With claude I might hit my 5 hour limit more quickly, but it doesnt leave me empty handed for the better part of 5 days! A weekly limit, on top of a 5 hour limit, is a bit too limiting and assinine. This is bad business.
I want to add another angle to this request.
The issue is not only that 5-hour and weekly limits may be inconvenient. The deeper problem is the mismatch between monthly billing and window-based entitlement.
Codex is sold as part of a monthly Plus/Pro subscription, but actual usage appears to be enforced through shorter windows such as 5-hour and weekly limits. If unused included usage does not roll over, project-based users may pay for a full month but only receive practical access to the remaining weekly/window allowance when they actually use Codex heavily.
This affects Plus and Pro users alike. Even if Pro provides higher multipliers, those multipliers appear to increase each usage window; they do not create a monthly bank of included usage.
I support the request for a monthly pool, or at least partial rollover within the billing month, unused-to-credits conversion, burst/project-based plans, or end-of-cycle grace usage.
After having upgraded to the 100$ new pro plan, let me say this. The
context limits you get at 20$ are no where near 1/5th of the 100$ plan.
It's more like, 1/25th of the 100$ plan. I could use my entire 5hr context
window on as little as two prompts. Under the new plan I can execute a
prompt that ends up running twice as long in fast mode and use less than
10% of my 5-hour context window on the $100 plan. There's no comparison
April's 2x context window on the $20 plan felt fair. May's supposedly half
of that context window feels more like a sixth. This hourly context crap is
not transparent in terms of billing. Not even remotely. And it is a
transparent attempt by open AI to attempt to force people into higher value
plans to compensate for the loss of subscriber ship due to their government
contracts.
This company has no qualms about who it will provide service to as long as
they are getting paid. Which shows a serious lack of morals and ethics
granted there's not a whole lot of that incorporations today anyway and the
Dodge versus Ford motor company case of 1919 isn't helping either. If you
don't know that case you should look it up.
But barring any further political commentary on my part I would love to see
a much more transparent method of billing and accounting that is actually
fair. Not everyone can afford $100 a month subscription plan to a language
model. I'm using it for work so it's much easier for me to justify and
acquire. But even at that it's still a tough pill to swallow financially.
But the clear obvious differences in amount beyond what is stated in the
sales info makes me question a lot more.
It's not going to matter what quality of model open AI produces when
everyone leaves and cancels their accounts because of the lack of morality
of the company running the models.
Do better.
On Sat, Apr 25, 2026, 01:31 feedback17 @.***> wrote:
Actually after examining the help docs, I see that the launch of the 100$
pro plan you get 10x codex usage instead of 5x -- just for the month of
May. What happened then do you make another $125 new plan and offer 2X
limits on that for the first month?
This pricing structure really is unsustainable. And people are going to get
sick of the bait and switch. I know I'm already halfway there.
On Sun, Apr 26, 2026, 02:05 Celeste Weingartner <
@.***> wrote: