Codex Usage Limits

[RayBytes]

Is the limits for the Plus tier in the ChatGPT app the same as the limits in codex? (3000 Thinking/week)

[Zamaroht]

I'd definitely also appreciate some clarification on the codex-cli limits for the Plus plan

[Dv-Dn]

ChatGpt Plus (GPT-5 Medium)

• Hit the weekly cap after ~6–7 full sessions.
• One “session” = fully using the 5-hour cap.
• In practice, that 5-hour cap is reached in ~2 hours of active coding.
• Net result: 2 days × 3 sessions/day, then the weekly cap triggers.

[shanecowherd]

Now I'm wondering what that cap would be if we use the api pricing? Would it make more sense to have Plus and fall back to API or to use Pro

[brainit-consulting]

Hi @shanecowherd , This is exactly what I was wondering, so I made myself two .bat files that automatically switches to the (pay as you go) API key mode when i hit the usage limit and back to the Plusplan + GPT login mode once they let me 'out of jail' and my free usage resets. Happy to show how it works if you want.

[erenersahin]

@brainit-consulting can you share some kinda gist of this?

[Kare-Udon]

📊 Token Usage
  • Input: 1491682
  • Output: 6819
  • Total: 1498501

ChatGPT Plus, one conversation, limit triggered.
I am also curious about the clarification on limits.

[Kare-Udon]

Well here comes a new record

📊 Token Usage
  • Input: 15927
  • Output: 140
  • Total: 16067

16K and being limited.
I may somehow pissed sam off.

[shanecowherd]

On Pro, last night I coded and coded, totally solved my problem, when I was done I saw:

• Input: 2249742 (+ 35030528 cached)
• Output: 209846
• Total: 2459588

Never hit a limit

[noelboss]

Coded 1.5 days: 19% of the weekly limit left 🥲 - going back to claude max…

[Aaron-Savron]

@noelboss exactly, came to codex expecting better quality speeds and Models. Got a bunch of fucking slop dude. OpenAI is such a scam

[dlukt]

it's way too low.
it didn't even finish a small service, now I have to wait 5h.
that's bullshit.

it did 6 tasks and then stopped mid task 7 of 9.

btw those 7 tasks were base skeleton only, no implementation.

teams subscription

[kaushalForge]

absolutely man the token finishes way faster i regret purchasing it in token terms, but the output it gives is good

[Nikollaych]

I'd really like to have a more clear info on how limits reset. It seems that I've hit weekly limit, and I have no idea when it will reset (Plus subscription).
Also, it feels that limits were getting gradually lower throughout the week.
First and second session were really generous on inputs/outputs, but then it felt even smaller than Claude code limits with pro subscription (can you even imagine it?)

[nuno5645]

i felt the same this week. My codebase was already large so i don't know what could it be

[mattyj-stack]

How large are we talking? Mine 300k in my monorepo and unless I want to throw money at Sam I am going back to manual coding lol

[smraikai]

Would be nice to have a warning like how Claude Code does...

"Approaching Usage Limit"

or something like that.

[ethantphillips]

They have added a warning at 75%.

[Symbai]

I reached the limit and I didnt got a warning

[AnkanMisra]

too bad

[Aaron-Savron]

@smraikai Just proves how much more Anthropic thinks about us to be honest, The fruits of labor are showing. Anthropic has a major Moral advantage and thats scary given the fact OpenAI works with the government. The circle of dumb shit. if money was the root of evil Open AI is the beast itself

[FreddyE1982]

I am totally puzzeled. I got Plus. It seems that limits in the web codex and codex cli are totally seperate. I have never run into the limit in web codex. I just now ran into the limit in codex cli, but web codex still accepts prompts .

[guywilsonjr]

Codex web counts as chat gpt usage not openai tokens like codex cli

[adenaufal]

+1 on what @guywilsonjr said
also i reached 5hr cap just now and went unusable if switching to codex web. however when i only using web, it feels unlimited as long as not using codex locally

[COWI-MI]

Codex Cloud should not be free anymore per doc https://developers.openai.com/codex/pricing

[adenaufal]

aw dang it :'( so its not free anymore

[dlukt]

Limits are better but still too low.
Seems like the Teams limit is now around 2M tokens.
It should be increased by 500k IMO

[dlukt]

Here we go again.
18:23 day 2, barely 7 hours of usage with breaks.
Yesterday started around 2pm ended around 2am. 19h of usage for 1 week.
That's not even 1 work week (40h) worth of usage.

[Sotiris-Bekiaris]

I just hit the usage limits with Plus and I get this message.

🖐  You've hit your usage limit. Upgrade to Pro (https://openai.com/chatgpt/pricing) or try again in 3 days 13 hours 6 minutes.

Usually it was try again in 5 hours... Is there a new usage limit?

[Kare-Udon]

Well, you may hit the weekly limitation.

[dlukt]

I hit the weekly about every 2 days, about 12h per day usage. Have to use a friend's account to continue (not rich). But even combined that's not enough to cover 7x8h days.
If OpenAI believes we need a day's rest they're wrong! :D
But yeah, the limits are too low and I read many posts on X that people with a Pro subscription (230€/m) have the same limits, roughly 2 days until weekly

[fxpayback]

I just hit the usage limits with Plus and I get this message.

🖐  You've hit your usage limit. Upgrade to Pro (https://openai.com/chatgpt/pricing) or try again in 3 days 13 hours 6 minutes.

Usually it was try again in 5 hours... Is there a new usage limit?

Literally the same, never hit any 5 hours limit, then hit the weekly limit straightly....

[Sotiris-Bekiaris]

Exactly! The funny thing is that in previous weeks I used it much more. I was even hitting the 5-hour limits. But this week I only used it once on Monday, and then suddenly I hit the weekly limit. I was also thinking maybe there are monthly limits too, but then again it’s strange that it says it refreshes in 3 days.

The limits and usage should definitely be explained in more detail. Unless they aren’t fixed and just get adjusted depending on how much you use it.

[dlukt]

What they're doing is borderline criminal.
The agent is working on something, fixing an issue, then in the middle of it you hit the usage limit.
Your project is left in a non-operational state. And you're supposed to wait, in my case 5 days 15 hours.
A Pro subscription doesn't resolve the issue, because the limits are the same as Teams (now Business).
That leaves the project non-operational, and you can't continue to work on it on your own, without the agent.
You're pretty much forced to create a new account and buy a Plus subscription.
As low as the limits are right now, that needs to happen every 2 days.
So you need at least 3, if not 4 or 5 accounts that you have to rotate in order to be able to work.

[wbdb]

Upvote here: #2992 & #2733

Compared to Claude, OpenAI already offers quite a lot for the money. But of course, some people don't need the 3,000 thinking requests in ChatGPT. So why can't they use it via Codex? It's legitimate to question this. In the end, a middle ground will have to be found.

[bfulop]

Last week I hit the 5h limit once, and the weekly limit on Sunday evening (had to wait 11 hours).

Tokens do cost money and we can do things on our side to reduce usage ("cognitive load" for the LLMs):

• restart often, document progress in AGENTS.md
• hints in AGENTS.md so the agents have to read less code
• clean separation of concerns: agents only have to understand 2 files as opposed to 20
• work in subdirectories
• plan with high reasoning, implement with low reasoning

Of course it depends on the project, but I'm trying to chunk my tasks in a way to have faster responses which usually means less tokens usage.

[dlukt]

• restart often: why? maybe for micro tasks, but having a big picture helps. Otherwise the goal is missing and you waste tokens because it always has to restart initializing the context
• AGENTS.md: I'm not too familiar. I thought it's for general guidelines, not progress/changelog
• SOC: That is not always possible and how to enforce that rule anyway?
• implement with low reasoning? But that's where high reasoning is important, maybe I misunderstand? Code quality needs to be good, for that reasoning needs to be high, else you'll waste time on iterations.

If you have to work to accomodate the AI to save tokens, why are you using AI at all then? It's there so help you reach your goal/vision, not the other way around.

And Currently 73€ is not $20 (re:linked post). 1 teams + 2 plus should last about 5 1/2 days. But I work every day. API cost is ridiculous. $1250 for 1M tokens. I'm not rich, but I'm working on it, and that's why I have to spend to gain velocity. Of course it can't be free, but this is predatory. It's on them to reduce their costs and also not destory our environment while expanding and improving and operating.

I have now skipped git commits, even if the summary is so nice.
I'm letting it work on the "hard"/tedious parts. Writing tests, auth rules etc.
But because the code it produces is not very readable, manual editing is hard to do after it has done major changes, even if you tell it to K.I.S.S. and DRY and build re-usable things.
I have become cost-aware, but it's still not enough.
And it always lures you to spend more.
And it does certain tasks, which are not required, and when you tell it to not do those, that costs tokens too, and not few.
So limits must be increased or the models and the interaction improved.
But why would they do that when they can upsell their products?

Edit: It's $1.25 (why are they writing it like $1.2500?) for 1M tokens
https://openai.com/api/pricing/

[bfulop]

Yeah even following my own rules, and being extremely careful, I hit the Teams weekly limit Wednesday evening...
And I thought that I was consuming less tokens this week than last!

It would be nice to see how much usage we have left per day/per month.

[whoschek]

I'm on the Plus plan and just started using Codex VSCode, first session. Worked fine for half a day, then I got the 5h limit and it asked me to wait for 30 mins. Fine with me, no problem, I have some other things to do anyway.

So far so good. But one day later I hit another limit (this is only the second time I've hit any kind of limit), this time it asks me to wait for 5 days! 5 days, you gotta be kidding! Not a few hours, but 5 days!

A tool quitting all of a sudden, out of the blue, for that long is crossing a red line for me. Totally unacceptable. I wouldn't mind waiting for a few hours, but waiting for days renders this tool totally unworkable in any practical setting. 5h limits are ok, but the weekly limits are nuts. Please fix the limits ASAP!

[wbdb]

Just use the API. The 5-hour limit corresponds to approximately €2-3 in API costs.

[bfulop]

Exactly: when I ran out of limits, I used the API and consumed about €2 per day.

[BlockViz]

They want to upsell you their Pro plan. Plus is for getting you hooked, then rail you in with the pro plan. I'm currently rotating 3 subscriptions, it's not as smooth as an uninterrupted session would be, but I feed it the last response and tell it to continue from there. Not optimal but also not paying 230€ but 73. Still a lot of money. The worst is that codex doesn't warn you about reaching some limit. And I think that's by design. They know it will leave your codebase in an unusuable state. I hate that it works. But that essentially Mafia methods. "Nice codebase you have here, you wouldn't want something happening to it"

How do you rotate? I purchased a new Plus subscription. Logged out everywhere and logged in with the new account. Still has the chat history in VS Code and also the usage error. I even uninstalled Codex and reinstalled it in VS Code. Still the same. There seems to be some issue with the session of the old account not being cleared.

[dlukt]

I'm logged in with my main account.
I copy ~/.codex/auth.json to ~/.codex/auth.json.main
Remove auth.json
Login with the other account.
copy ~/.codex/auth.json to ~/.codex/auth.json.second
Remove auth.json
Login with the third account.
copy ~/.codex/auth.json to ~/.codex/auth.json.third
Now there are

auth.json.main
auth.json.second
auth.json.third

create scripts in /usr/local/sbin/
codex-main
codex-second
codex-third

/usr/local/sbin/codex-main

#!/bin/env bash
cp ~/.codex/auth.json.main ~/.codex/auth.json

/usr/local/sbin/codex-second

#!/bin/env bash
cp ~/.codex/auth.json.second ~/.codex/auth.json

/usr/local/sbin/codex-third

#!/bin/env bash
cp ~/.codex/auth.json.third ~/.codex/auth.json

Before starting codex I run e.g. codex-main
Since 0.36.0 you can use codex resume to resume the last ression, but you have to check approval and model

[BlockViz]

Ah, thanks for the reply. I think you are using CLI, not the VS Code Codex Extension. It seems the credentials are stored differently there. I do not even have the .codex folder... :-/ Maybe I need to switch to CLI.

[Thelastpoet]

Let me stick with CC

[dpolivaev]

I opened a feature request addressing some of the discussed issues in a different way:
#3366

The proposal is to let Codex CLI continue a session after hitting plan limits by allowing re-authentication — entering a new API key or logging in with another account — instead of terminating. This would prevent losing context or leaving code in an inconsistent state.

Vote for it if you like it.

[guywilsonjr]

I mean can't you just pay and set up an account with 5 dollars to allow finishing the last prompt?

[quantuminformation]

maybe they all going into price fixing

[zj21st]

Hi bro, I found the reason. This is a product bug — the server was sending the remaining usage to the plugin, but the plugin mistakenly interpreted it as used amount.
Just update the plugin to the preview version 0.5.36, and it’ll be fixed — in that version, it’s clearly labeled as “remaining usage.”

[magnus-m]

new model gpt-5-codex-mini should allow 4 times more tokens.
Also got 50% more tokens with chatgpt subscriptions.

https://community.openai.com/t/codex-updates-mini-model-higher-limits-priority-processing/1365540

[quantuminformation]

oh it just reset to 100% weekly, it said it would on 25th

[grashupfer99]

For me it said the next weekly reset was on 26th, the next day it did reset itself and changed the date to Nov 30th.

[dlukt]

https://x.com/thsottiaux/status/1992370994028388670

[quantuminformation]

github santa doing his remote rounds early

…

On Sun, 23 Nov 2025 at 12:19, Alex Hunt ***@***.***> wrote: For me it said the next weekly reset was on 26th, the next day it did reset itself and changed the date to Nov 30th. — Reply to this email directly, view it on GitHub <#2251 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABU35VUL2V6FA4W5K4H4CD36GJ4HAVCNFSM6AAAAACDYMECQCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMBVGIYDSOI> . You are receiving this because you commented.Message ID: ***@***.***>

[nerdwave-nick]

Wtf is even the point if this stuff keeps hitting limits. What am I paying for. A broken product? If my toaster stopped working after 2 slices of bread, I'd throw it out. Maybe time to do the same with this AI slop.

[dlukt]

They reintroduced the enforced limits and lying about it.
In plain text, when you hit limits, it ends at once instead of ending after the run was completed, leaving your codebase in a broken state again. Likely to upsell their Pro offer

[etraut-openai]

@dlukt, to my knowledge, we have not changed anything on the client or the service in this regard. If you are seeing behavior that doesn't match what you expect, please use the /feedback slash command to upload the session and create a bug report.

[dlukt]

The abrupt ending is definitely back and you're lying about it

[k2xl]

I purchased for 40$ additional credits after hitting weekly limit but the limit did not lift! I got 1000 "code review credits" which i do not want.

Also unable to process refund for this

[etraut-openai]

The limits that you see in the UI are your subscription limits not the available credits, which apply only when you hit the normal subscription usage limits. It's done this way because on some plan types (notably, Business plans), purchased credits are pooled among all users in the plan. It wouldn't make sense for them to appear for individual users.

If you purchased additional code review credits, you should be able to continue doing code reviews even if you've hit your subscription limit.

[k2xl]

Fortunately OpenAi processed a refund for me.

But, @etraut-openai it was confusing for me and let me explain why. First I get this:

I visit the website

I see this. I hit Add credits assuming that it will resolving the error. It probably shouldn't ask me to Add credits if it won't solve the problem I have.

[quantuminformation]

limits are great now thx, i can keep 2 clients happy with just gpt +

[dlukt]

... they're temporarily doubled until April you...
And the regression still exists that it stops working mid execution
etraut is still a terrible innovation blocker

[cobycloud]

at 47% weekly usage, 53% of 5 hour rate limit, yet I'm still being rate limited...

nothing better than shadow rate limiting my rate limit's rate limit

I mean.... scream to me your infrastructure is unstably oversubscribed, without saying out loud, you know... go ahead and just bake that policy right on top of the RAG of your "support agent" and call it a day, am I right?

It takes some seriously bad character to deliver half and then serve up the rest as a toast note.
I think about it in terms of like... going out and buying a meal at a restaurant.

Could you imagine ordering a tomahawk steak and then they come out with tapas and toast?

[ljubogdan]

I noticed that my token usage seems to be significantly higher than before. The same tasks and prompts now appear to consume several times more tokens than they did previously. I did not change my workflow or prompt size. Has anything changed recently in how tokens are counted or how responses are generated (e.g. reasoning tokens, system prompts, or model behavior)?

[olsgo]

codex team, i'm not sure what's going, but the limits situation is a joke at the moment. it's totally unpredictable. in less than a day after the reset and i already ran out for the week (reset only takes place on the 18th), on the pro plan. do you guys know what could be happening?

[rldyourmnd]

I had a weekend and didn't do anything for a couple of days (with 95% of limits). I thought I'd come back and use up my weekly limits for the last two days, and then start over. My limits were reset, so I came back from vacation and spent 75% of my weekly subscription in the first two days. It was kind of a joke. Maybe you could reimburse me somehow? Well, if you have something reset, try to keep people's limits, but give them a voucher to reset their weekly limits - that would be much fairer.. :(

[lhluengo]

alguien a probado el pro? cual es el limite de esa suscripción? en la pagina dice ilimitado

[e6o]

Great thread. To answer @shanecowherd's question about API vs subscription — here's the math based on current API pricing:

GPT-4o API rates (per 1M tokens):

• Input: $2.50 / Output: $10.00
• Cached input: $1.25 (50% discount)

Example: Your typical Codex session
@Kare-Udon's usage above: ~1.5M input + 7K output tokens

• API cost: ~$4.45 per conversation
• ChatGPT Plus ($20/mo) with ~6-7 sessions before cap = ~$3-3.30/session equivalent
• So at moderate usage, Plus is actually competitive with raw API

The break-even shifts once you:

1. Hit weekly caps and need to wait 5h windows
2. Run multiple agents/threads in parallel
3. Use higher models (o3, o4-mini) where API gets expensive fast

For heavy Codex CLI users who also use other LLM APIs (Claude, Gemini, etc.), one approach that's been working well is using a routing proxy that automatically picks the cheapest capable model per task. I built InferCut for this — it's a drop-in proxy (same OpenAI-compatible API, one line to switch) that routes requests to the most cost-effective model. Typical savings are 30-50% on API bills because not every task needs GPT-4o.

For people hitting caps here and considering the API route, the key optimization is model tiering: simple tasks (file renames, formatting) on cheaper models, complex tasks (architecture, debugging) on premium models. Most people send everything to the same model and overpay.

There's a detailed cost comparison table here if anyone wants to see per-model pricing across providers: https://gist.github.com/e6o/23db9719f8d9c173234ec7471e044a36

[SuccessMoneySparkle]

useful tips, but before june 4 reset 100-150m tokens were 10-14% of $100 weekly allowance. After reset 150m day was 40% of my weekly allowance asking reddit and discord multiple people noticed the same issue. For I used same mode, same setting, limit got cut overnight. One user shown his token usage on same plan 280million token used, costing him 48% of his daily usage. That puts our $100 usage to about 300-500million. While some people have 1.5billion tokens on the same exact plan. I used to get 2.5billion with 2x promo and 1.2 after promo ended with $100. Before that I used $20 which had 245million per week using the same 5.5 Xhigh.