This link was here for people such as yourself.
I’ve been using codex for about 7 hours today and spent about 9% of my weekly on one pro account. I’m also extremely specific in my prompt scope and don’t let the agent do a whole lot of thinking.
So if your issue is legitimate, the link there is for you.
I do have a business account and plus account under the same logins. It was done this way so I could keep my personal stuff seperated from business. However I do use codex on both accounts because I do some stuff for work which I keep to their account and my stuff to my own account. However recently when ive hit the 5 hour window, I have switched to personal to finish off what I was doing.
I have also wipped out ~/.codex and even ~/.vscode thinking maybe something simular to what you mentioned, but more about possible corruption but this made no difference.
This also explains so much, as probbaly my average requests is around 100k and ive been finding ~47% drop in the 5 hour limit just over that single request. With 1-2 requests, the whole 5 hour window is wipped out. Its roughly 30-60 seconds of writting a prompt, will completly take out the whole 5 hours. No one in their right mind is going to pay for a subscription if they only can use it for 1 minute every 5 hours which currently is roughly all you can do.
Also 100k tokens for codex as you clearly pointed out is actually pritty normal. In my case, its mosty working on a couple of repos, they are not overly large in size, but do have about 200 lines of AGENTS rules which does incress the input token count. But without these rules, codex keeps doing the same silly mistakes, so they are mandatory.
The abnormally fast limit drain, in my opinion, is directly related to the fact that the system is valuing tokens much higher than it should. In other words, the issue is not the amount of usage itself, but how that usage is being priced.
If you look into it, this becomes quite obvious: the current subscription economics appear significantly less favorable than using the API, and in some cases even less favorable than premium models like Claude Opus. These inconsistencies are easy to spot. Anyone who has encountered this can check their own data and see that they are effectively paying far more than expected for token usage.
In this context, explaining everything purely by the end of the promotional period or changes in pricing doesn’t really hold up. Those factors alone don’t explain such a large discrepancy.
At the same time, it seems that not all users are affected. Otherwise, this would already be a much bigger issue that couldn’t be ignored. Right now, it appears to impact only a relatively small portion of users.
This leads to a possible hypothesis: this could be related to a new system or logic introduced by OpenAI to combat multi-account usage. From a business perspective, this would make sense — it’s not beneficial for the company if users bypass limits by maintaining multiple $20 accounts.
With that in mind, I’d like to ask those experiencing this issue:
have you used multiple accounts or combinations of subscriptions (e.g., business + plus, plus + pro, multiple business accounts, etc.) while working with Codex?
Any feedback would help clarify whether this is a bug or an intentional system behavior.
I need to push back on this, because your post suggests you are not speaking from the perspective of someone who actually uses Codex in any meaningful way.
Please ask your team to test Codex properly using the same types of accounts many of us have, and to do so in real-world tools such as the CLI and VS Code. They should try normal day-to-day tasks: simple git commits, small refactors, and other lightweight coding workflows. That would give a far more accurate picture of what users are actually experiencing.
At this point, almost everyone in this thread is reporting the same pattern: within roughly 20 minutes of genuine usage, what is supposed to be five hours of allowance is effectively gone. From what I have seen and personally experienced, around 200k input tokens appears to consume roughly five hours. For a coding model, that is extremely difficult to justify. I am only counting input here, because output is usually minimal, often around 100 to 200 tokens per request.
Once normal weekday usage resumes and more people are back to work, I expect this thread to grow quickly. At that point, you may be forced to re-evaluate these decisions, because users will not continue tolerating a product that burns through usage this aggressively. If the only meaningful feedback mechanism available is for people to leave, then that is exactly what many will do.
I believe theres more to it, its Easter weekend (just gone) and many people leave and don’t touch work over this period. That comes to a close tommrow and many will return back to work. I predict within the next 24-48 hours, this issue will be widespread.
The thing is, while looking into this, I noticed something important: there’s essentially a split happening on the forums between two groups of users.
The first group — those for whom everything is working as expected.
The second — those experiencing abnormally fast limit drain.
The problem is that the first group often assumes the second group is just using their limits inefficiently. They find it hard to believe that limits can actually be consumed as quickly as described.
Or the explanation is reduced to: “this is just the new pricing, you have to accept it.”
As a result, the second group ends up being unheard — and that’s us.
Yes, we are not the majority, since not everyone is affected. But there is one pattern I’ve noticed: a large portion (50%+, maybe more) of affected users seem to have something in common — multiple subscriptions, particularly combinations like business + plus.
I will be using an API key for now. business account is not really worth it. its hard to transition projects and internal documents to the chatgpt, since codex cannot reference chatgpt conversations seamlessly.
I think the split in the chatgpt website and codex is too big. they should have never made codex locked behind a secondary payment system without offering a codex focused plan that limits the chatgpt website product’s usage.
It is one or the other, and they over did it and made the chatgpt website product more usable than it has any right to be. simple as that.
My friend, you are driving for answers that are out in the open. Business and Enterprise, and seemingly at least some plus users have been moved to the new rate cards. It says it right on the website, and was confirmed by the openAi rep here. It’s coming for the rest of plus and pro users. Free lunch is over, and it’s across the board, at least for the big US based companies. Anthropic, Cursor, windsurf, OpenAi, and many more are all pretty much at same place for pricing now, and they’re done subsidizing cost. If fuckin sucks, but just means we have to keep adapting, and not allow ourselves to be vendor locked.
bro tell me pls where is inefficiently using in this prompt:
“replace in the \src\assets\icons folder parameter “fill” to “fill=“#ffffff”” in each of 17 SVG files.“
and after this he thinking 3 minutes and used 2% of weekly limit for 17 short strings
it is unusable
Sorry for the friendly fire — I didn’t finish reading and thought you were on the other side
You can’t reduce rates for existing users TENFOLD.
If it’s not bug, no one should ever buy annual subscription. Codex is now useless on Business
My hypothesis has been confirmed!
The hypothesis was as follows: the anomalous rate limit consumption that some users in this thread are experiencing is most likely related to attempts to bypass usage limits. For example, when limits run out on one account, the user switches to another account with a paid Business subscription and continues working.
Previously OpenAI didn’t track this, and the scheme worked fine — you could pay for several Business accounts and switch between them as limits were exhausted. But it appears that in recent days OpenAI has started cracking down: this kind of activity is now being tracked, and affected accounts receive a hidden restriction that makes each request abnormally expensive.
To test the hypothesis, I created a clean account, paid for a subscription, and started working on the same projects and tasks — the limits are being consumed correctly. This confirms that my previous accounts had a hidden restriction applied, causing a single request to consume 10–50% of the five-hour limit.
This is not 100% confirmation — we haven’t received an official response from OpenAI yet. But it solved my problem and allowed me to continue working normally with Codex. Hopefully this is useful to someone.
What kind of account ? Plus?
Yes, I purchased the “Plus” subscription on the new account I created.
Yes. No new rate (still count the number of messages) on this. The other one is a Plus too?
I just started work today like usual. I must admit that today, for some reason, things seem “back to normal” again?
The past few days I’ve been hitting my 5h limits after like 3 minor tasks or questions. Something as simple as “hi” would cost me 3% of my 5h limits if not more.
Today, I’ve been working with it for about 2 hours straight and I’m only down to 60%, which seems quite normal and usable again.
Did something change? I didn’t get an update notice or anything.
I’m still not fully convinced by your hypothesis, and here’s why.
After exhausting both my Business account and my personal account, I asked our MD if I could temporarily use his account so I could finish some work. He has never used Codex in his life and cannot code. He only does very light tasks in the GUI. Even on his account, I experienced the same extreme usage behaviour.
That strongly suggests this is not simply down to account history, user habits, or someone “using it wrong.”
One thing we are all aware of is this note:
“As of April 2, 2026, we’ve updated Codex pricing to align with API token usage instead of per-message pricing. This change applies to new and existing ChatGPT Business and new ChatGPT Enterprise plans. Customers on existing Plus, Pro, and Enterprise/Edu plans continue on the legacy rate card for now, with migration to the new rates coming in the next few weeks.”
Because of that, I suspect part of what you are seeing is the difference between Business and Plus pricing, which would absolutely create a noticeable change in perceived usage. I noticed the same thing myself and mentioned it as well.
That said, this thread still does not reflect the true scale of the issue. I checked Reddit out of curiosity to see whether others were seeing the same behaviour, and it is full of similar reports. In fact, it is far more active there than this forum.
I also think OpenAI may be underestimating how quickly competitors will move to capture the users it loses. That is simply how business works. If customers feel they are no longer getting fair value, competitors will step in immediately and take advantage of that.
I do not particularly like saying this, but large tech companies rarely seem to pay proper attention until customers hit them where it hurts most, which is revenue.
Based on what I’ve seen, I am going to recommend that my boss cancels all Business accounts that are due for renewal next week, and I will also be advising clients to do the same.
Update
This could also explain many things.
This is really putting my hypothesis as a big question of truthfulness.
Sorry, its also possible they are doing tweaks behind the scenes and the new acct has them and rest don’t. We trully don’t know. Only time will tell.
I’m happy to report that my hypothesis was most likely incorrect. It was probably just a coincidence, a fluke.
Because I just logged back into my account, and everything started working properly again. I asked a couple of questions, and my limit didn’t even drop from 100% to 99%. This suggests that everything is working just as it did before. And perhaps OpenAI has already fixed the bug that was there. I’m incredibly happy that I can return to my old accounts and continue working on my applications.