Usage dropping too quickly
Resolved 💬 325 comments Opened Mar 5, 2026 by sibonyves Closed Mar 12, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
Feedback thread ID thread ID 019c8e57-2892-7be1-9304-fc8c29bce3af
I didn't copy my feedback text in it but it was detailed enough. Main idea is that usage limit dropped way too quickly, I think I was downgraded from 2X limits back to 1X limits, and started consuming credits :(
325 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Having similar issue right now with gpt-5.3-codex-spark. Regular gpt-5.3-codex seems to be consuming limits as usually, while with spark I've spent about 15% of weekly limits (Pro) in 10 minutes.
Same for me! Consumed like 40% of my weekly usage today, using GPT-5.3 Codex, both Codex App and Codex CLI, this feels unreal!
Seeing this as well - Pro account, consumed 10% in a couple of hours. Past entire month the 2x quota was working perfectly
gpt-5.3-codex highWoke up to 27% missing after spending the last several hours before that only having used a couple
E: Due to the number of issues I am experiencing I have stopped all of my codex processes and the usage is still being consumed, lost another 3% between then and now
I'm seeing a drop of 1% on every prompt. Had two small changes, ran a
/reviewconsumed 1% on weekly.i upgraded to the $200 pro plan last night and have used 10% of my weekly limit within 24 hours...
Thanks for reporting the problem. I thought we had fixed this problem and put it to bed, but it sounds like there's still some issue. Apologies. We'll try to get to the bottom of it quickly.
To help us out, could I get user IDs from a few of you who are seeing this? That will help us accelerate our investigation. To get your user ID, go to https://chatgpt.com/codex, click on icon in the upper right corner to display the profile menu, then click on the top menu item. This will copy your user ID to the clipboard. It should look like "user-<random string>". Post that below so we can directly look up your account.
user-jYCnlbS06oGZr3keY5J5TG78
E: v +1, the numbers go back and forth wildly, each TUI session reporting different values, and the website also being different. I vastly preferred the case where it felt like I didn't have to worry about this and could simply do the work I wanted to. This amount of usage is substantially accelerated past "2x not being applied", the raw session logs that the TUI app writes are indicating 15x consumption
Appreciate your reply. My user ID:
user-zRdWbpZQTFZAoOIvqgi7iTtqFor your info:
I'm open to more question in case you need them (or provide logs or whatever).
Thanks.
@etraut-openai
user-XGVKRizj4f3SBIIUqva0hmSG- mine's drained toweekly 88%with just a couple more prompts. Essentially 12% in just a few hours.<img width="1874" height="758" alt="Image" src="https://github.com/user-attachments/assets/c8b81a22-1bb4-47ef-aff3-0431b5079d32" />
You can see my weekly usage per day for the entire week, compared to just today. And today was perhaps a very slow day, did not use codex much at all.
user-WO80XRMuQXS5ncJJ6FLJ2Zlp
Thanks, I think that's enough user IDs for now. We're investigating.
Extra feedback that might help, not sure though. I saw a strange issue on Codex CLI again just some moments ago. Using GPT-5.4 xhigh my weekly usage said 97%, then suddenly after auto compaction it said 78% weekly. It sure shocked the hell out of me. However I checked the Codex usage on the website and it said 97% weekly. So I assume it's a glitch in the response that Codex CLI is receiving regarding my account's quota usages. It seems to currently be flipping back and forth periodically between 97% and 77% weekly on the TUI. It's like it's reading someone else's quota at times perhaps? Maybe this is why some people are seeing their quotas burn faster too, if it's also depleting another account.
Not sure if it helps but my User ID is user-fxCyi5bYo7JQMuevvS9mTzli
Any update on this? I would love to continue trying to use codex but combined with the permissions issues I'd need to sit here like a drinking bird pressing approve every single time it touches a file.
Have you tried GPT 5.4 to fix it I heard its great
It sounds like there are potentially two different issues here:
Do I have that right?
Yes.
if my spark-specific case would be useful, user-dKr49OiKhzHyeI81FkW3Sb1U
@etraut-openai I'm not sure if 'wild swings' is the correct description but yes I did notice that my weekly quota suddenly disappears from the status line (sometimes while it works) and then re-appears later. I haven't paid much attention to it but yes at one point I saw it say 97% and then later I saw 88% but I can't be certain because my status line shows both the current context remaining as well as the weekly quota remaining.
Overall I've under utilized codex today and I'm already at
weekly 82%within half a day when my quota officialy was reset this morning. With the 2x quota I'm usually 82% after 48 hours roughly.That’s correct, but in my case I’ve experienced both. Massive usage consumption and crazy swings in usage reporting.
Edit: plus my limit didn't get reset even though it was announced.
I've also had some swings in what the CLI was showing (the usage page online was steady), sometimes it would go back up to 90% then back down to 80%.
No, I am seeing both issues. I experienced a 13% drop after my new week started. Then it got reset yesterday. Now I'm somehow down to 70%. I have done nowhere near that much work.
The extreme usage seems to have stopped for now, but the weekly usage is still swinging. Each session is reporting a different value +/- 1%, most of the time it doesn't even appear in the status bar at all.
Tallying the token usage amounts from my logs gives me a massive number that makes no sense. Some amounts are duplicated exactly.
I typically have to use swarms of codexes to use my weekly usage, and despite using it as I usually do, I went through 40% of my weekly usage in a single day. I wasn't using it any more than I usually do........... Whatever the fix was for this was a regression for me. Please give me my credits back.
@bijanmoney, are you using the new gpt-5.4 model today, by any chance? It is more expensive than the gpt-5.3-codex model and will consume credits faster.
@etraut-openai No, I have been using the same workflow/model 5.3 codex xhigh for weeks now, nothing has changed on my end. I didn't even know 5.4 was available.
Likewise. For comparison, the type of things I would have to do to see either amount of usage in that timespan are the irresponsible things I'd try when the week is almost over and I'm lighting tokens on fire, 6 simultaneous xhigh's with sub agents type stuff. But the extent of my usage are two sessions I've been working from all of the prior week, 85M total tokens (2.3B cached) and another that's currently busy so I can't check but I at least remember has ~4.5B cached. I'm vaguely concerned that you are unable to immediately verify what is happening now that you have my User ID. Is there anything else I can provide to help?
@etraut-openai I am also concerned because this has been an issue now for over 24 hours and you haven't acknowledged it. At this rate I will be out of credits in 2 days which has never happened before, and normally I also throw in a bunch of extra work, now I'm having to be very cautious with what I do and it's still burning credits like crazy as compared to 48 hours ago.
@bijanmoney, this is one of the top ten issues we're actively tracking and working on. We're juggling a lot today because we launched the Windows Codex app yesterday and the gpt-5.4 model today. I understand your frustration and appreciate your patience as we work through these issues.
@etraut-openai Much appreciated. And here's my userid: user-REO6tl215YaeMAmQct4eX4Ss
also running into this issue. i think the usage started draining much faster yesterday even with 5.3-codex. i have not seen it fluctuate. i have also been using 5.4 since it came out in standard service tier. my user id is
user-gwZMDLcfFcwY8OlkizaGkqe2<img width="1463" height="838" alt="Image" src="https://github.com/user-attachments/assets/1258d2a0-12bb-4bff-89b5-5f34b8755d9b" />
I upload this picture where you can clearly see how my usage fluctuated in like 1 or 1.5 hours. Went from like 34% available, up to 1% remaining. Then, it's 27% remaining again (in both CLI and website).
Edit: after 12 minutes (since I posted the last image), take a look at my weekly usage available. During this time, I have not used Codex.
<img width="1403" height="788" alt="Image" src="https://github.com/user-attachments/assets/0d28a202-a1f0-4eed-8915-afe4133cb9ff" />
Can confirm I notice this behavior at least 2 weeks back where there was a bug that was claimed to be fixed.
I was affected by one of the 9% of claimed PRO & PLUS users. So I am monitoring whether anything changed. I did swap to 5.4.
If usage limits were indeed applied then I'd expect the rate of usage drop at least a bit with the same workflow?
<img width="2194" height="836" alt="Image" src="https://github.com/user-attachments/assets/9381a84d-16b6-42a1-bdc6-ef83055649b5" />
User id
user-EwwvXRbDXkpKEnwW9TeVbKGTsame here. my codex usage did not change that much, at least from what it feels like. last week i suddenly for the first time managed to deplet my weekly usage limit (usually i have 40% left). now this week im already at 68% after just 1.5 days and i didnt even do that much. has to buy credits for the first time haha
my userid: user-ynqUgf40LvP2F3x7eWAGOMRL
I mostly use Codex through VS Code with almost only gpt-5.3-codex (medium thinking), and I was among the ~9% of users affected by the previous bug.
Since the reset, I also have the impression that usage consumption is faster than usual.
user-8R4S5miJVczqPQEiBKOHDuVU
In my config.toml I enabled the following features:
Could this possibly be related?
Good luck with the investigation, and thanks for all your work.
please check mine as well: user-jPTXF7p71zITKWcY7ix6ovjK
user-vV3XhezUrPWzLBc4F3sjbF9C
<img width="124" height="239" alt="Image" src="https://github.com/user-attachments/assets/9e266589-5841-40a3-bdaa-e269f0bc558f" />
<img width="1181" height="179" alt="Image" src="https://github.com/user-attachments/assets/d01bae59-8d37-454a-8171-d7107cdf34d2" />
This is clearly broken.
Same here unfortunately. My user ID is: user-MMETTtNhsSapsl61oIYfGrpu
I'm noticing the same issue. I'm on the Pro plan. I went through almost 70% of my usage in one day. I usually end the week with at least 10% or more remaining, and I don't feel I had an abnormal workflow that day in comparison to other days. My limits just reset on the 4th at 3:30 PM.
User ID:
user-lkqg4raupjouznjF0npCN8cKBased on an analysis I've done of the logs and the token usage, I believe this is largely related to the
fork_contextfeature onspawn_agent.---
What happened
A single session consumed 1.16B tokens (~50% of my weekly limit). The session started as "run e2e tests" (tests passed in 40s), then continued with follow-up tasks across 50+ user messages. The session wasn't active the entire time — I just had it open and returned to it throughout the day.
The model spawned 22 subagents, 12 with
fork_context=true. The ones spawned withfork_context=trueconsumed 70-220M tokens each. The ones withfork_context=falsewere all under 2M.Token breakdown
| Agent | fork_context | Tokens | Duration |
|-------|-------------|--------|----------|
| Parent thread | — | 230M | — |
| Dewey (worker) | true | 219M | 5 min |
| Avicenna (explorer) | true | 157M | 6 min |
| Arendt (explorer) | true | 142M | 5 min |
| Pasteur (explorer) | true | 78M | 5 hrs |
| Boyle (explorer) | true | 78M | 4.8 hrs |
| Aquinas (explorer) | true | 77M | 6 min |
| Raman (explorer) | true | 77M | 5 min |
| Nash (explorer) | true | 69M | 5 min |
| Godel (worker) | true | 23M | 13 min |
| 12 others | false | <2M each | — |
| Total | | 1.16B | |
The model's fork_context decisions
The first 3 spawns correctly used
fork_context=false. Starting at spawn #4, the model began usingfork_context=truefor tasks that don't appear to need full session history:The forked subagents inherited the full session state (79 user messages, thousands of tool results). Dewey's rollout shows 84 turns, 16 compactions, and 1,061 tool calls in 5 minutes.
Config at the time
max_threads = 8,max_depth = 2Session details
019cbfdd-a496-7b31-b017-a0c1e1467051If it's at all helpful, mine was with the subagents feature completely disabled.
Yes, no sub-agents etc here. Mine are being consumed by simply running
/reviewI went through 40% of my usage in a day (probably intermittent usage over 7-8 hour perio).. Using GPT-4.5, this is insane given we're supposed to be on 2x until April. I have never went more than 10-15% in a single day with 6-8 hours of work.
I've spent almost my entire week's rate limit on Pro now, in just a few hours, so I really do hope this is just a bug.
I am also seeing significant usage. My user ID is user-i5JKeTgOvYaEyyHqtwZEjEsA
Subject: Codex 5-hour / weekly limits are depleting unusually fast
Hello,
I’m reporting what looks like abnormal Codex usage-limit depletion after the recent update.
ChatGPT User ID: user-ZaTPWLAaN2r0dI5KghE35Y29
Issue:
My Codex 5-hour limit and weekly limit now decrease much faster than before, even during normal usage. I am not using Fast Mode, so this does not appear to be explained by higher-speed consumption.
What I’m observing:
Why this seems incorrect:
Previously, similar sessions consumed noticeably less of the 5-hour and weekly limits. After the update, the same type of work appears to consume a disproportionately large share of the quota.
Request:
Please check whether:
If useful, I can provide approximate timestamps of recent sessions where the drop was unusually large.
Thanks.
please check mine as well: user-0FsnAGdb6b320VNiPNGvqLNo
its been a week since i am facing this issue, mentioned it in #13186 last time i ended up using entire quota within day, and it didn't reset as they announced. Now that i get weekly reset, it depleted 50% in few prompts.
I think I’m affected by this too, on the high-consumption side.
My dashboard is now at 35% weekly remaining, but my actual recent usage still does not look unusually high compared with my own normal usage.
From my local usage history:
2026-03-04: 30.83M2026-03-05: 90.07M2026-03-06: 24.56M so farSo my post-reset usage is only about 49% of my previous average, yet my weekly quota is dropping much faster than usual.
I also accounted for
gpt-5.4pricing on2026-03-06. Repricing that day asgpt-5.3-codexwould only reduce the estimated cost by about $3.75, so that does not explain the overall mismatch.That’s why this still looks wrong on my account.
Codex user ID:
user-FiFz006Brg4lzM7gHBGzOEoIEdit: since I posted this comment, my dashboard dropped from 35% remaining to 24% remaining.
In the same interval,
ccusageonly increased by about:That additional usage is heavily cache-read dominated, which makes the speed of the weekly quota drop look even more suspicious on my account.
Another data point from my local session logs: I did not use sub-agents on March 6, 2026.
This comment was written by Codex under my supervision.
Reporting same issue here, usage is abnormal.
EDIT 07/03/26 due to giving resets only to user id posts: user-I8aSs3aU5EXwixiLJh2BOmL8
@etraut-openai What should be done while the issue is being fixed is to restore the limits on all affected accounts. There are many of us who weren’t even given the limit reset you did a few days ago. I personally haven’t done anything since midweek because I still have 4 days left until my limit resets.
@J3m5 btw last time I checked, ccusage had a bug for counting codex usage that wrongly doubled some numbers, so I suggest you to ask codex itself to count its own history
For those of you who are using gpt-5.4, an increased usage rate is expected in this case. It's a more expensive model. See this documentation for details.
If you are using gpt-5.4, please do not add to this thread because it's muddying the waters as we try to find the root cause of the problem.
The problem is affecting gpt-5.4 for me, with 5h/weekly limits remaining dropping far faster than expected - beyond the expected increased usage rate as documented above. user-WDSJnoTbHYk5cckGAhWfPDgy
I have the same problem: when I use GPT-5.4 my limits drop absolutely unreasonably fast literally, my weekly limit is decreasing at the same rate as my 5-hour limit. My ID is
user-RoQRZ7w2lIsuhSXhtaC18aRqPRO subThis can't be explained by an increase in the model's price, since its price rose by only about 11%~, not by 4–7 times - that's absurd.
Same with 5.4-high, limits drop about 2x as fast as they should, despite not using /fast or context bigger than 256k
GPT-5.4 usage limits are only 27% lower compared to 5.3-Codex, me and other users are seeing 40-60% drops, so wdym?
What about gtp 5.2-codex? Is this also affected? I was suffering of this
issue last week and switched to 5.1-mini to save some credits. Now I have a
complex task I would like to use 5.2 for but I'm afraid it will eat up my
credits without finishing the task.
Il ven 6 mar 2026, 18:04 insanowsky @.***> ha scritto:
@etraut-openai this isn't very clear; are we saying GPT 5.4 is expected to take 2x the tokens (and so effectively our 2x quota giveaway would result in 1x ?) And come April 2nd, we'll see GPT 5.4 burn 2x faster, with
/fastburning at 4x ?Having said that, this problem began with gpt-5.3-codex, only switched to using 5.4 last night, seeing the same burn rate.
Update: the limits in documentation do not tally with what I'm seeing; it started with 5.3 and only after a _fix_ was announced to people complaining earlier (at which point I somehow was not affected), and with 5.4 I'm seeing no difference. It's as if the 2x limits have been removed on my account.
There's something wrong with the usage and the limits. I'm affected. Claude has this sort of problems and a lot of people left Claude. Now it's Codex. What is going on?
Edit: I'm seeing this using 5.4, NOT the fast mode
I've been exclusively using GPT 5.2 high, only just started exploring GPT 5.4 a few minutes ago. But I burnt through 54% of my weekly usage limit in just 36 hours on the $200 Pro plan with average use. Often I wasn't even hitting 50% of the usage limit at all in 7 days.
I’m seeing something similar on a ChatGPT Business / team account in Codex CLI (
v0.111.0).Current usage readout on my side:
5h limit:47% leftweekly limit:33% leftcontext window:25% left(198K used / 258K)gpt-5.4 (reasoning xhigh, summaries auto)What feels off is that usage seems to drain very quickly relative to normal local CLI work, even though this is a Business seat.
From the current OpenAI Help Center article on Codex + ChatGPT plans, my understanding is:
2xCodex rate-limit uplift across paid plansIf that is still accurate, my Business seat does not seem to reflect it consistently.
It would help if the team could clarify:
5h+ weekly caps,2xuplift is actually applied to Business seats,reasoning xhighsessions are currently consuming quota faster than users expect, or if this is a usage-reporting bug.Reference:
https://help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan
I can confirm that the usage is WAY off. I was using 10% a day, and now I used 31% IN AN HOUR.
If this is becoming like Claude, which you have to wait 5 days to code again, then I can tell you this is the wrong direction.
I just want to confirm. I never reached 5h limit, now I hit it all day long. Was using 5.3-codex without any problem, only weekly limit. I am now down to less than 10% weekly and my usage wasn't changed.
Just before a weekend man what a bummer.
user-y3YtfijmyL1UgfuclpSvQFzKSame with me. Been using 5.3-codex high for weeks now with no issues whatsoever for either the 5 hour limit or the weekly limit and now after a day or two my weekly limit has 29% left and resets march 11th somehow now? Also today I only have asked for one or two prompts and my 5 hour limit is down to 65% somehow. Also have not been using fast model at alluser-OEnOb45qRl3xjUAgbTZyxoCl
Same issue for me. Mainly using 5.3, then switched to 5.4 and limit disappeared with minimal use.
Edited to add: For me, the issue mainly seems to affect the weekly limit rather than the daily limit, i.e my daily limit has declined 2% in 2hrs, but the weekly limit has dropped by many multiples of that over the same period.
From yesterday my weekly limit almost 60% completed with just normal use, before 6th march everything was good, but suddenly my weekly limit resetted on 6th then 60% completed
I haven’t even touched version 5.4. I’m still trying to use 5.3-codex without it removing 10% of the weekly limit over a trivial task. And I started noticing this on Sunday, as I already mentioned in other issues.
We're continuing to investigate the problem. Sorry for the delay — we were overwhelmed with several concurrent incidents.
We have a couple of theories about this issue, and we're pursuing them all. We just made some server-side adjustments related to one of our theories. If you are experiencing this problem and have a quick way to test usage (like a small prompt), please give it a try and see if you are still seeing the issue.
Usage draining for me way too fast too on Pro plan, Using 5.4 xhigh, fast mode off, in a day it drained almost 40% of weekly usage, prior to 5.4 release the way I use codex I was able to drain this in 3-4 days of usage.
Are any of you who are affected using a client or harness other than Codex? We allow other harnesses to be used with our ChatGPT subscriptions. I ask this question because one of our theories, based on our logs, is related to non-Codex harnesses.
No, just windows codex app and codex cli
No, Codex Extension for VSCode, pre-release 26.5304.20706
& codex-cli 0.111.0
No, using Codex CLI 0.111
No, macOS codex app
user-jJB61U0yuwId4MkMAeyb2qVB
Affected as well, codex-cli 0.111.0 in VSCode Insiders on WSL. Consumed about ~50% of weekly usage in 2hr on pro plan with GPT-5.4
@etraut-openai purely codex CLI, latest alpha
codex-cli 0.112.0-alpha.4no other harness.A few of you have noted that the usage for the 5-hour window looks OK, but usage for the weekly (7-day) is suspect. That's an interesting observation and potential clue. Has anyone seen higher-than-expected 5-hour usage consumption as well?
@etraut-openai me - the 5 hour usage I see is high as well. It seems to burn equally quickly.
I can't say for sure, but the 5-hour limit seems roughly reasonable. However, the weekly limit flies by, in my experience, at about the same pace as the 5-hour one.
Both 5hour and weekly, they never drained so fast.
In my spark-specific case there was no misalignment between 5h and weekly usage draining. They were both decreasing with regular ratio.
In my case, I’d say I only used Codex for the past days. I’ve noticed my weekly usage runs out much faster than my 5-hours limit (which I don’t think I’ve ever hit, unlike my weekly limit). Like it drops my weekly usage by 30%, and it equally drains my weekly by a same amount (I don’t really think that makes sense).
For me both are pretty high. Yesterday I burned through the 5h in 4h (had never happened to me before), and this consumed 30% of the weekly. Not sure if that's expected or not (I never really looked at the 5h before since I never reached it).
Based on what I'm hearing, both the 5-hour and 7-day usage are affected by this issue. Thanks for that confirmation.
Both are high for me too.
@etraut-openai Hi, would it be possible to reset our limits after the bug depleted them? We’d really appreciate your help. Thank you.
This problem had been solved two days ago and today it has just returned
This is exactly what I'm seeing - 93% 5hr usage, and 18% remaining. I've been using solely 5.4-high and 5.4-medium today - /fast for a while, though after seeing usage rapidly climb, switched off /fast, and still observing a quick increase in usage.
A quick smoke test: ~40 tokens input prompt, 43k tokens used total, GPT-5.4-high, MacOS codex UI client. My weekly usage remaining dropped 19->18%. My 5hr usage remaining dropped 95->93%.
Agreed bug literally took my whole week tokens 🥲
As I mentioned above, we made some server-side adjustments related to one of our theories. If you are experiencing this problem and have a quick way to test usage (like a small prompt), please give it a try and see if you are still seeing the issue.
Also, some of you reported in this thread that you were seeing the reported "usage remaining" values jump up and down. Are any of you still seeing that behavior? We're trying to determine if that's related to the higher-than-expected usage rates or if it's a separate issue.
@etraut-openai The usage is still jumping up and down. The fix seems to work for me, but it’s still dropping 10-20% faster than before. That said, this fix does let me use Codex normally again, and I still have 30% of my 7-day usage left. Also, the usage in the status line still disappears sometimes.
I don't think it's fixed for me. I haven't measured tokens, but over a few small prompts the weekly dropped 3% against 5% for 5hr window, so extrapolating would suggest one 5hr window usage would take out 60% of the weekly limit (which doesn't seem right).
A few of you have mentioned that you're using the experimental multi-agent feature. Are all of you using this? That could explain increased usage relative to having that feature disabled.
I'm not
I also mentioned in the previous thread that the issue also happens on the free tier, not just paid like you said.
@etraut-openai i have it enabled, but i'm not using it
Are you guys using websockets v1, websockets v2 or no websockets at all?
Hello just marking sure you didn't miss this reply earlier. I believe this could a the root cause of the issue for people who have multi-agent enabled.
---
What happened
A single session consumed 1.16B tokens (~50% of my weekly limit).
The model spawned 22 subagents, 12 with
fork_context=true. The ones spawned withfork_context=trueconsumed 70-220M tokens each. The ones withfork_context=falsewere all under 2M.Token breakdown
| Agent | fork_context | Tokens | Duration |
|-------|-------------|--------|----------|
| Parent thread | — | 230M | — |
| Dewey (worker) | true | 219M | 5 min |
| Avicenna (explorer) | true | 157M | 6 min |
| Arendt (explorer) | true | 142M | 5 min |
| Pasteur (explorer) | true | 78M | 5 hrs |
| Boyle (explorer) | true | 78M | 4.8 hrs |
| Aquinas (explorer) | true | 77M | 6 min |
| Raman (explorer) | true | 77M | 5 min |
| Nash (explorer) | true | 69M | 5 min |
| Godel (worker) | true | 23M | 13 min |
| 12 others | false | <2M each | — |
| Total | | 1.16B | |
The model's fork_context decisions
The first 3 spawns correctly used
fork_context=false. Starting at spawn #4, the model began usingfork_context=truefor tasks that don't appear to need full session history:The forked subagents inherited the full session state (79 user messages, thousands of tool results). Dewey's rollout shows 84 turns, 16 compactions, and 1,061 tool calls in 5 minutes.
Config at the time
max_threads = 8,max_depth = 2Session details
019cbfdd-a496-7b31-b017-a0c1e1467051I’m not using any new feature. Neither multi-agent stuff nor GPT-5.4.
I’m using whatever the Codex add-on in Visual Studio uses by default. Either way, it also happens with OpenCode.
Using the experimental websockets v2
Just tried the same setup with spark again. With 2 parallel agents in just 10 minutes I got down by ~60% of 5h limit and ~20% of weekly limit. Slightly slower pace than I reported here the first time but it still is absurd at Pro account.
user-ZXVCiuGykBnl6P5vFEID4y5TI think we should focus on reporting only “normal” use cases, without complicating factors like multi-agent usage. I’m saying this because they clearly don’t know what’s going on, and they’ve already made several attempts to deflect by implying that if we were using GPT‑5.4 or multi-agents, that would explain everything.
I only use the Codex Visual Studio addon, and out of curiosity I also installed OpenCode to see if the same thing would happen — and indeed, it did on both platforms equally. I think you can rule out that theory
Sorry if I made it sound more complex than it is. No subagents were used, and I just ran two instances of Codex at once. That's it. 20% of weekly limit consumed by two instances means 10% of weekly limit consumed by one, which is unusually much.
Tried 5.3-codex again, and it seems to be consuming limits at an expected pace. Might really be something about websockets, since only spark uses them by default as far as I understand?
Regarding the multi-agent + WebSocket v2, I haven’t changed those settings over the past few weeks, yet I’ve observed a very recent increase in consumption. So either you’ve changed or upgraded your algorithm related to these options, and they’re now being used differently in a way that consumes more.
I'll keep the same settings for the next few hours, and starting tomorrow I'll remove multi_agent to test whether there’s a noticeable difference.
I have not changed this setting. I was never previously burning usage at the absurd rate I am today
If I were to reflect what change, I started using gpt-5.4
To me this seems like the caveat, but shouldn't it be more token efficient?
Using 5.4 medium in both CLI & extension and realized it's eating my tokens like candy. Down 40+% in one day. On Pro subscription. Never has it been this bad. Context window is default & compaction is default & not running fast mode. Something's off
Monitoring usage on a bigger prompt: daily usage dropped by 26% against 9% for weekly, giving a ratio of about 35%, i.e. three 5hr sessions would use the weekly limit. Compared to the earlier ratio of 60%, that seems like an improvement, but don't seem indicative of the double usage allowance being in effect?
@kali113, you mentioned that you're seeing the usage stats jump up and down. Could you provide your user ID?
Brief stats over the last hour or so, if it helps:
This is for ~80% GPT5.4-high, 20% GPT5.4-medium
5hr usage limit left % / Weekly usage limit left %
93 / 18
90 / 17
88 / 17
85 / 16
84 / 15
83 / 15
82 / 15
81 / 15
[Switched to GPT5.4-xhigh]
81 / 14
80 / 14
[Switched to GPT5.4-xhigh fast]
78 / 14
77 / 13
75 / 13
..
36 / 1
@moogman, thanks for the details. Can you post your user ID? To get your user ID, go to https://chatgpt.com/codex, click on icon in the upper right corner to display the profile menu, then click on the top menu item. This will copy your user ID to the clipboard. It should look like "user-". Post that below so we can directly look up your account.
user-Qbut9FVm3kBlgO9tffbTxYnC !
I only use codex cli
Also - this issue started EXACTLY when the usage was reset to 100%. It was not a problem until that moment. I monitor my weekly usage very closely, it's always visible to me as I work. I've only ever considered that I wasn't using enough of my credits literally until that exact moment that it reset to 100% this last time.
@etraut-openai I checked, my limits are being used up at a similar rate to moogman's.
GPT-5.3-codex, High/Extra High
Codex extension for VSCode, pre-release Version 26.5304.20706
codex-cli 0.111.0
Windows 11
user-fXm3qn3kHwvvZECtVFyKZZ7NI hope this helps you.
Just chiming in as another datapoint. I managed to go through my entire plus plan in one day. user-RY7JwID71Bqqj3QHgK3xOm8H
Codex CLI and Desktop were used in this time. No other oauth. MacOS
For anyone that has multi-agent enabled and has noticed this issue, I suggest adding this to your root Agents.md in the meantime until it's fixed. Based upon my analysis, sub-agents are being created with
fork_context=truein situations where they shouldn't be. That causes a massive increase in token consumption.- Never use fork_context=true when spawning subagents.user-98S8pihNQFWOg5yQvWdnqE9E
GPT-5.4 xhigh and high
codex-cli v0.111.0
Mac 15.7.4
As we go into the weekend, I wanted to give you all a quick update. We spent all day poring over logs to try to understand the issues reported in this thread. Our investigation was complicated by the fact that during the past week, we enabled a number of features that could potentially affect usage metering (websockets, gpt-5.4, fast mode, etc.). We needed to review each one and eliminate them as potential causes.
We uncovered one issue that is affecting some (and perhaps many) of you who have posted to this thread. It has to do with usage rate accounting across regions (data centers). Most Codex users will be routed to the same region (the closest one) for every call to the model. However, some users will be on the boundary between two regions, and their requests can get routed to multiple regions. We try to keep usage rate accounting in sync between regions, but they can get out of sync briefly. That explains why some of you were seeing usage percentages jump up and down. It may also explain why some of you saw sudden drops in usage percentage. We are working on a mitigation to further reduce this sync latency, but this is a fix that we don't want to rush because it could negatively impact all Codex users if we get it wrong. We are pretty confident that this sync problem is affecting a very small percentage of our users, so we're going to take the time to get this right.
In addition to diagnosing the sync issue, we spent a lot of time looking for other potential reasons why users might be experiencing higher usage consumption. We triple-checked that the issue that we experienced earlier this week has been fixed, and we're convinced it is. We also ran down a bunch of other theories and disproved them one by one. We haven't yet found any evidence that there is any incorrect usage accounting.
We'll continue our investigation next week. As thanks to everyone who supplied their user IDs and helped provide clues for us in this bug report thread, I've manually reset your usage back to 100% so you can enjoy Codex over the weekend.
If you are experiencing usage consumption rates that seem higher than in the past, here are some things to keep in mind:
Thank you very much for your very professional reply!
I will try using codex over the weekend and will report back :)
edit: For information I am located near Geneva Switzerland, I don't know where your servers are located but maybe I am switching between Zurich/Paris/London?
Didn't use any fancy thing, don't know what websockets are, issue was before 5.4 was launched, as soon as the rate limits were reset on Wednesday evening/Thursday morning, same workflow as before.
Thanks for the reset. I had actually been wondering about the issues with accounting going out of sync; I do frequently have cases where I have several sessions running at once and see different numbers for remaining usage in
/statusoutput in different sessions. Are you currently leaning more toward this being an issue where usage is getting counted multiple times when the out of sync accounting state is being reconciled, or is it looking more like the 2x limits aren't applying to an account for another reason altogether, or maybe something with the endpoints in one region not being able to properly determine the context window size or fast mode status of sessions in a different region? I can definitely see why debugging metered usage ticking at some multiple of the intended rate would be a huge pain in the ass when there's a double usage promo and multiple new options that each individually increase the burn rate all being rolled out simultaneously.user-sMFfXWYr9SyKZC59VEvl1sQE
Codex 5.4 high and xhigh. Suddenly went through 30 percent in a day though it seems to be better now.
user-wfLYrLFd1Yu6Cq7wMquDZSBU
Thanks for the reset.
Thank you! I really appreciate that.
Were you guys able to figure out why sub-agents were seemingly randomly getting spawned with the context fork parameter, even in situations where it seems like they shouldn't have been?
If there was a distinct pattern, it would make more sense, but it just seemed really inconsistent. I'm kind of concerned that will continue to happen in the future. I can provide more info if you'd like as well.
I've added guidance in my agent's MD to try to prevent it, but that doesn't seem exactly ideal.
user-10PG4YWFUqJVm6ztXyKRiQPv
Appreciate you guys trying to figure this out, especially on a Friday! Thanks for the reset.
user-0cKp93NZgrnaFQVDM4ulHtz9
Providing me user ID so you can check if I was affected.
ChatGPT Pro, WIndows 11. Could be because I use a VPN? Not sure. But surely saw the highest consumption of my rate limits today.
Props to the team. Have a good weekend. Thanks for your timely response and tackling the problem!
user-MkUurS8X16OR8tZYVmxJHevF
Still seem to be having high usage.
Usage limits still seems to be dropping quite fast. 5% left of weekly limit now. I am using 5.4, but have used only 5.2 high/xhigh before, not 5.3-codex. With high usage before I only barely managed to spend the full rate limit (2x) in a week, but now it took barely a day. Now I realize 5.4 is faster than 5.2, so that might be partially to blame, but even so, it still seems to be burning fast. I do have subagents enabled, but the faster usage seems to have started at the launch of 5.4 or 0.111.0. I do not have /fast enabled.
user-d6qXI5UVUNsKXKlHxGT4rZ9L
user-suuLtzaIFlpzxj7MMxMvSVVV
I’ve also run into this issue: my usage limit started dropping much faster than usual. Previously I was able to use Codex for an entire week, and I was impressed that, unlike Claude, the Codex CLI really felt suitable for serious agent-based development. But after switching to gpt-5.4, I noticed how my weekly limit was exhausted within a single day. I’d appreciate it if you could take a look at my account as well.
Overall, it seems to me that this issue might be more widespread than you think, since there are still quite a few users on Reddit who are expecting a global reset of limits due to unjustifiably high usage consumption.
sharing my userID here for further checking
user-VEIdJezFp4YTLB7HqBxZSquY
@etraut-openai Thank you for reviewing our cases. However, my usage still hasn't been reset. The website shows 27% available, but when I send a prompt from Codex, it says I’ve reached the limit. The previous reset wasn't applied to my account either (and though this one was supposed to be a manual reset). I'm mentioning this in case it helps identify what’s going on.
I'm one of the early users who sent their user ID for you to review (
user-zRdWbpZQTFZAoOIvqgi7iTtq)<img width="1584" height="765" alt="Image" src="https://github.com/user-attachments/assets/c7b5f825-d672-4c2e-b57f-2eb96b73eb87" />
user-4hHWeZcEUiHNcJvhrUIkcu0G
I'm also experiencing unusually high usage depletion rate.
I'm having this high consumption rates since last week.
My user id is user-ZJKeqW2iMU39Y050NZWrfLtU
Thanks for resetting it over the weekend.
Cheers
-----------------------------------------
Antonio Bulgheroni <http://www.google.com/profiles/antonio.bulgheroni>, PhD
On Sat, Mar 7, 2026 at 9:55 AM Kaid Wong @.***> wrote:
I also have seen an extreme high consumption rate in the last 3 days with no change in my workflow, if you could look into it, my user id user-aFT2XFCzEENYPqtyiEDbOaaw
The first day I noticed higher usage I was using gpt-5.3-codex, then switched to the new model, but the spike was even before gpt-5.4.
Thanks for looking into it.
I've had the same experience as others in this thread. Before the usage reset I was already using subagents with gpt 5.3 codex high. After the reset (and before I switched to gpt 5.4) I noticed a drastic increase in the rate my usage dropped. gpt 5.4 accelerates that even further which is expected, but I was seeing the increase usage with 5.3 codex too.
user-cpR5OWKW23aI1HFMMjOXZeys
Hi guys where can I find my user id so I can apply for the reset? My entire pro weekly limit went down 90% in a single day. I'm looking everywhere for this ID but its not in my account settings or anywhere I can find.
@etraut-openai I noticed that on the last day, March 6, my usage seemed lower than on other days in the Codex usage graph at https://chatgpt.com/codex/settings/usage, even though I had already used more than 30% that day. I’m also seeing the usage go up and down unexpectedly on the Codex status line, and sometimes it even disappears. I didn’t check the usage on the web very often.
Could you please reset my usage? My user ID is user-VDuca11J9MCHEkyz5RTFGSoa. I’m located in Valencia, Spain.
<img width="689" height="249" alt="Image" src="https://github.com/user-attachments/assets/8a0bfa5e-f4d3-46fe-8c6d-aea9c8fd5a02" />
My codex used 40% of my pro account in a single day, this is crazy.
user-xXAv1ikAnQM1xo6JneDgfbDh someone please check my account
Please reset my credit usage too, i'm down to 0% now. User ID:
user-Wag7D1Uwx0Te02PY1KfpOoc3i really appreciate the reset especially on a friday. after the usage reset today, i have disabled multi-agent and responses websocket v2 features (i had explicitly set them to true before) and now the usage limits are draining at a much more reasonable pace even with multiple concurrent threads.
We did not provide the user-id because you said you had enough, we are also affected and asking for same treatment please:
user-I8aSs3aU5EXwixiLJh2BOmL8
Where is everyone finding those user ids?
Go here https://chatgpt.com/codex/ top right corner click on image then click on image again it autocopies to your clipboard.
I think this is my user id, unless codex is hallucinating it: user-W6Yry1phH347lov6mU8g8TfA
Please reset my limit i had to stop yesterday after burning 91& of my weekly limit in a single day.
Are you guys for real? So you ask for IDs, 1 hour later you say that you already have enough IDs and then you just reset the people that sent IDs? So the people that are affected and didn't send their IDs because you said you had enough are left out to dry. Some of us already came from the other thread where you said you fixed the issue and I immediately said the explanation made no sense since the usage was going down by more than double, so it can't be related to the x2 promo not being applied.
I started noticing this on GPT 5.3 Codex, then you made some changes and my limits barely went down at all and then you did more changes and closed the other thread and the increased consumption restarted. I'm not using subagents / fast mode and the issue is not limited to GPT 5.4.
If this is indeed normal, then once the x2 promo ends, our limits will be basically useless.
ID: user-VJ396CjqKfjSGQ5jR9hYMGFj
My weekly limit will last a few hours post April 2nd.
Same here, been using it yesterday on PRO account and somehow I got an usage of 50%, this makes no sense.
ID: user-2jzCMG2jPD4VuweI2ttp0WWS
user-grEQqq7dokHb56wxlYbBeUWO
Same issues, Pro plan here... usage limits burning fast yesterday(from 60% weekly to 7% now) and i've used only gpt-5.3-codex xhigh + gpt-5.4 just to orchestrate(no fast or 1M context).
I agree with @calves07, when this promo of x2 will end, CODEX will be unusable, because it is burning too fast tokens - this is not sustainable.
i ****ing used up 80% of 2x pro usage with gpt 5.4 xhigh in less than 14 hours !!!
WTF
I already did a comment about my weekly usage yesterday before you guys mentioned anything about resetting the weekly tokens.
Now I read that you require us to post our ID to fix it. Here it is.
ID: user-ARpTy5d0ZgjgKzFSwRc4h0wB
To put this into perspective, what I normally use within a day is 10%; with heavy work, what CODEX does now is use 35% of week within an hour or two.
where do you guys get the user id from ???
user-x97zg5W5XrBKINyNO5QviDmf
Somethings wrong with my usage
@waml click your profile icon, and click email
"me too" doesn't help here, it just makes tracking even worse. You can 👍 the issue to note this instead.
Secondly, please add some sort of useful information with your post - even a screenshot of your 30 day usage page, or a list of a few of your last threads with in+out tokens and which model you were using. This might help the team triage.
Just for your information: I also have swinging values of consumed credit with ups and downs. This could be due to geographical boundaries. I work from Germany, on a development center in Italy and the headquarter is in Belgium. So depending from where my VPN is re-routing the traffic, it can appear from different countries.
Ask codex xhigh it's one of the hardest things to find any other way. I fed it the link to this thread because it didn't understand what ID I meant, saying it doesn't exist, etc. They make it really easy...🙄
I’m having the same issue
@etraut-openai I shared the wrong userid in my post, The business account in question is linked to this user ID: user-bJGntHa632NL8RPprQzE9eiz
Can you reset my weekly limits please?
My user ID is: user-gOB1mqZ5mabbtlpT73XBbcB7
I’m seeing the same issue. In my usage breakdown for the 30-day view the daily usage in the week ending March 7, 2026 is much higher than prior weeks, mostly under Extension/local usage rather than Cloud. The dashboard only shows relative percentages/bars, so I can’t verify exact token totals or per-action consumption from the UI. It looks like my effective rate/consumption changed during this week even though my workflow did not change materially.
Model used was 5.4, usage was before starting, but suddenly jumped by high amount.
My user id : user-TpfZNYJfzmTjJqBJqLt0AnEP
Yesterday i barely used GPT 5.4 xHigh, and i am on the 30% limit, which is weird, considering i also have 2x until April....
My user id is : user-IZ4kQatkVNtvvyBIcZs26kqK
Since I started using GPT 5.4, it burnt 65% of Pro Credits in about 2 1/2 days.
<img width="1216" height="350" alt="Image" src="https://github.com/user-attachments/assets/404e86e7-3797-4cb5-b592-7ef45786b91b" />
My id : user-166HxLTEZoj1XxPF0US5i1zq
Thanks for the analyze !
@etraut-openai Thank you for the amazing work you and your team is doing every day!
user-pdeFisbIXyiOlgmPFNlBehQb
<img width="2262" height="610" alt="Image" src="https://github.com/user-attachments/assets/eeff5bd5-a47e-40ff-bd00-238ef02a066d" />
I thought the high usage was just because of the Windows app release, but today the usage seems to be at a much more reasonable level compared to the last few days.
user-lgkxYam8a20K1TTK2JlWfQ46
<img width="1040" height="225" alt="Image" src="https://github.com/user-attachments/assets/d6b5cb06-f29c-4c83-aeec-5d1c949cde01" />
Thank you for the amazing work. Is this the same situation that your team mentioned earlier?
Yesterday, about 50% of my Pro Plan usage was consumed within just 2–3 hours.
Today, I am trying to run as few subagents as possible, and it qualitatively feels like the situation has improved during the process.
Heya, what's going on here? I was just using MEDIUM, not even high or extra high and burned through 63% of my weekly in a day. Is this the new normal now?
My ID is jorakay
My credits burned upp in under 24 hours from just doing max 15 requests. While this would be a week's usage before, I had some additional credits I bought as a buffer for when I would go over my weekly limit those all got spent as well before I noticed how fast I was burning credits. My user ID is Dem0sthenes
user-ASRBruydQSXstUlHF5Cz3LCw
<img width="2012" height="352" alt="Image" src="https://github.com/user-attachments/assets/4a54f4a1-cb65-4c8f-a946-4271c02cd85c" />
<img width="580" height="418" alt="Image" src="https://github.com/user-attachments/assets/d08bedbc-8826-49d1-a8c0-8dcace87733e" />
This seems wrong.
On Mar 6-7, 2026, the usage table shows about
547Mtotalgpt-5.4tokens, but the weekly bar already dropped to82% leftwith11% in deficit.Previously I was using around
300Mto500Mtokens per day and the weekly percentage did not drop this fast. Right now it feels likegpt-5.4is being counted much more aggressively than before, or the weekly meter is bugged.I see similiar problem. already activated the second max account and working on it now. user-kGYX5YlSkqYhmKDlqDSQabhc this is the new one, i hope the error is fixed.
user-PDlMnHQbtl6VNDEdOq3Om3mD this one was draining extreme fast. looked like the weekly was dropping half the speed of the 5hour limit.
I'm seeing the same issue. My user ID is user-nZ80Ttjkz5rXi8PkOn8YPvR9
Same here. My user ID: user-sCs2M9xqM5hTnXe5oNRNJf3T
@etraut-openai Codex limits aren’t syncing properly, and the usage indicator in the status line occasionally disappears. See the attached pictures.
<img width="1920" height="110" alt="Image" src="https://github.com/user-attachments/assets/f2ae92d5-002d-4a80-8e57-d353a45a9ae3" />
<img width="644" height="223" alt="Image" src="https://github.com/user-attachments/assets/7ef1ab96-3eba-4af2-a796-74f1926bfa33" />
I have very abnormal usage these two days. User ID: user-6885OaLWRCCPEWeeye4CxGye
On https://chatgpt.com/codex/settings/usage, there is a "Other" category using much usage but I don't know what it is.
<img width="2048" height="761" alt="Image" src="https://github.com/user-attachments/assets/70f5d961-0d1f-4bda-92be-1d6484b0dec1" />
Same here: user-La8U1pDRI8jGWQn3flcIA8rd
Same here : user-sB9I37P7DLlfx2dZkpWeKWDy
Same here for my second account. : user-7jfhP1DTu6TXtdIiZaLvgDVp
The 5-hour and weekly limits are running out very quickly.
My weekly usage has also run down much faster than usual:
user-AfQOki8g6pHUlqcpCjdMUhJGSeeing the same issue here. My user ID: user-RFAlbiOKiKhqXLNRsHAUTyJG
!Image
Attaching my usage chart for the past 30 days. As you can see, on March 5, 2026, the Extension usage spiked to 100% of daily capacity, completely out of pattern compared to every other day. Usage was consistent and normal before and after that date, which makes this look like an isolated billing/metering event rather than actual usage on my end.
That chart always shows 100%, it's the share of each platform in the given day. So on that day, you only used the extension
user-8NyLSpYq37UflIGl9ipI8AkP
user-lcpUcIR4SYgflk89OUXDVXlD
two users, weekly usage dropping fast. on the user-lcpUcIR4SYgflk89OUXDVXlD, i woke up to 0% weekly usage remaining, and the othe rone is dropping fast.
as well as the 5hr usage limits are being consumed at a rate i've never seen before for both accounts. never experienced them in codex 5.2 and 5.3, despite workflow remaining the same and working on the same repo.
Hi, my weekly usage has also run down much faster than usual. Especially yesterday. user-0nOv6llo2UpRn4SX9XGJUM7s
I’ve also experienced my weekly usage going down really fast
user-J1Uk61YWOMSmQtWy48MTKx0j
Same here: user-exdRGIs8C56pmSpdIpC8lx6f
Mine draining real fast...
user-lbjHsSXPdHDsqLVyd02zjAXX
Mi weekly usage drained too fast too, empty in just two days, not normal at all
user-vlKY9sMmdQ3qUv2Ox6taAzX3
Same problem here, almost never use all weekly usage, this week it's empty in just two days
user-cFLmTzSFvUwiuPGpJtggUvmc
Great reset please. You will code nothing and be happy.
Same problem here! It's the first time i used all the weekly usage, it literally emptied in 3 days (in the past weeks I have never gone under the 40% of residual, but my daily use it's been literally the same as always)
user-RgkUgYZTD61UVyJ05B1NdfLh
Same Problem, My usage usually lasts for an entire week but its over even with my normal usage, I've a business account
My user id is this: user-4fun5QaVfEzkqscwlTxipHmL
my limits are being drained ridiculously fast... even though i'm using 5.4 high at most and 5.3 medium usually. 4% in just 15 minutes!
user-oEj3RuuWyOemWY0yOGndT6NP
Hi, Same problem. Burnt the weekly limit in just a day.. Pls help.. Thanks :)
user-c9CmeRF4xN9CqFCn4m58JLn8
Additionally to my comment here: https://github.com/openai/codex/issues/13568#issuecomment-4016341062
Codex with GPT-5.4 has been working for 4 hours straight now and only consumed 3%. It seems that something has in fact been fixed and that the original issue did affect me.
user-fmPfxpIk4bGh9CX7OW1A5Lvt
Since it consumes a lot of power, I turn off /fast and use it little by little.
[features]
fast_mode = false
Can you share your settings? Stuff like context, websockets, sub agents, fast mode, reasoning, etc
Somehow burned through 30% of my weekly limit on the first day of a 7-day period. Have never done that before even at these usage levels. user-RGFBH0OMypEAW5UWoaQTm2th
Note that in the usage chart my usage yesterday (Mar 6) was well below many other days where my usage did not contribute to a 30% weekly limit consumption. I don't think I could've even been using Codex as much as I have at the rate my limits were being consumed yesterday.
<img width="2358" height="746" alt="Image" src="https://github.com/user-attachments/assets/b769506a-9ef2-4528-901f-a5aaeae09ba2" />
Also having the same issue:
user-N4Jm887BFYwGzYnE8sgGUxqi
user-1TVKerujUjfTRBKWNfOlXymA
Also please, give us a $100 plan 😢
Hi, I have the same issue.
Yesterday about 50% of my weekly limit was consumed, abnormally high usage.
user-NC3823MdkNE9lAP2IA2iCVTa
User ID: user-Mm2qVha5FBaZHkkltnPX7E8f
Since March 4 experienced abnormally high usage (see screenshot) until I ran out early hours of March 7. I don't even use the fast mode or subagents. See screenshots:
7 day usage
<img width="1204" height="463" alt="Image" src="https://github.com/user-attachments/assets/d72a5822-aa8f-4d3e-9dd5-efab79d8a973" />
30 day usage
<img width="1206" height="428" alt="Image" src="https://github.com/user-attachments/assets/3441debe-e402-486c-a7c7-4307d89d8e88" />
user-q52cpEEcd8uXyPrBANXaq26U
Same issue here as everyone else. Down to 37% of weekly using Codex using only Medium, single threaded.
I’m having the same issue. Over the last week, I used up my limits in just 2 days, whereas previously they would last me about a week. My user ID is user-5HcW1OtGDeHujK0XzE1mOX9j. Please check my account and reset the limits if possible.
I definitely feel I got overcharged for my consumption yesterday. I usually dont even near my weekly limit, especially not in a single day. Also my credits were eaten alive.
user-FldIYhVAuPdWd5lfldLouv2b
user-5U8fX5H6gkuaGkC4ACJ0LF3u
Same issue here as everyone else. Down 40% of weekly using Codex cli, 5.3 on high
user_id: user-WKsP6c47dYur30DZo8yTGEMN
Weekly limits getting eat up at the same rate as 5hr window limit. pls help, thanks!!
Hi, I’m experiencing the same issue.
My weekly usage is being consumed much faster than usual even with normal usage patterns.
User ID: user-4yNuwuujQqEI4da2YjASOWvy
user: user-YVGujdhKLZrD1at2C8vY1Tjj
my limits get depleted almost 2x faster, i dont use context past 256 or fast mode and even after calculating gpt 5.4 lowered limits, its worse than it should be - gpt5.4 should cost me 27% more than 5.3codex and im seeing 2-3x more limits drained
Have somehow managed to blow through 75% of my Pro limits in a matter of days whereas previously I wouldn't approach that level of usage in an entire week. Have not used any faster processing or expanded context models. A little higher usage I could understand... but having to ration my usage after only two days seems fishy.
user-WYB6tLYOVPX6uNaUaxWNNMyj
User ID: user-OhVveO8UjR0PQ1H6ezwcHKaw
I have never used all of my weekly usage and used 80% of mine in two days this time. Something definitely is not right.
Editing to mention i am using default settings inside my IDE. Not using /fast or custom context size. I have previously used codex for >40 hours in a week without using my full usage.
!image
user-jXrOtC2l1RGSjt9xU9GHKRCK
Personally I know I was doing quite a bit of work, but my usage does seem higher this week so I figured I would toss you this in case it helps in your investigations if nothing else.
<img width="1243" height="433" alt="Image" src="https://github.com/user-attachments/assets/4f3dd8c2-54ff-4614-973c-daeeb41c9c4a" />
Same here - usage went by so much quicker than the past few weeks. I'm only a few days in an am almost out of tokens whereas before I never ran out. My usage for 2 days matches what previously would be the whole week.
user id: user-LRCM8OO62DUncNmS41lUQwX4
Same problem here, some relaxed use has used more than half of my weekly quota in a day or so, thats more than twice as fast as usual.
user-nCMbeXaQCHUBAfIXMcXXmPgL
user id: user-nacySpkGmwnIMeLOHSZldqdA
GPT-5.4 xhigh and high
codex-cli v0.111.0
Mac 15.7.4
You can look at March 5th as the anomalous date. Doing more work today to see if resolved, but that day was at around 5.5x normal. No behavior change. Almost at limit because of that day (of course).
Usage limit spiking unexpectedly
User ID: user-qHg3tes4WtTilBsz0O19Xyyh
Description:
Within a very short period of time, the usage counter dropped dramatically, without any corresponding workload on my end. The consumption jumped within seconds, as if multiple tasks were being billed simultaneously or a single operation was counted multiple times.
Steps to reproduce:
Open Codex and start a task
Observe the usage indicator before and after the task
Notice that the remaining quota drops by a disproportionately large amount relative to the actual work done
Expected behavior:
Usage should decrease gradually and proportionally to the actual number of operations performed.
Actual behavior:
The usage limit dropped sharply within a very short timeframe. The spike appears to be inconsistent with the actual workload, suggesting either a billing/accounting error or a race condition in usage tracking.
Additional context:
This aligns with the original report in this issue (thread ID 019c8e57-2892-7be1-9304-fc8c29bce3af). It is unclear whether this is related to a tier downgrade (2× → 1×) or a separate counting bug. No unusual tasks were running at the time of the spike.
user-PH6QsA7GXJVzUav8tkxG1Xei
Usage limit dropped way too quickly
usage limit seems excessively high considering recent usage compared to last week
User ID: user-LEIJkCk1xzqdqrgeqjszYGK4
comparing today's usage vs other days, I seem to already have consumed over 25% of my weekly limits despite having analogous usage over the last several days, of which consumption did not expend nearly this fast. I am now suspicious that there is an accounting error of actual vs reported usage.
Now I don't know what to trust because there is no user-facing visibility to keep the system accountable.
Same problem happened to me in the last three days. Just one hour coding spends me over 30% week limit. My usage limit may exhaust tomorrow at this speed, without limit refresh until Mar 12th. The abnormal usage spend can be clearly seen when compared with the usage record through the Feb.
<img width="1644" height="522" alt="Image" src="https://github.com/user-attachments/assets/75c52332-e986-41fa-a9c7-0bac1c0aa583" />
user ID: user-KjSgsCp8zI3xoR2ZEXDAoorN
user-DMVj9wG4wgnWABLqs3EWllEh
Abnormal usage limit consumption
hey, adding my +1 to this. been hitting the exact same token drain issue today and yesterday. looks like codex 5.4 is aggressively burning through tokens, specifically when running in "fast" mode.
it's hard to debug or approach this on my end since we don't have access to granular usage stats or telemetry, but the math is definitely off. my standard workflow just chewed through a typical 5-week token allocation in a matter of a few hours. normally, this exact same workload comfortably lasts me for several days of heavy, intensive coding.
i'm attaching a screenshot of my usage spike for context.
user id for the logs: user-UGXfvyOAuz3t4wPVFIlEJU8g
let me know if you need any specific prompt examples or local logs to help isolate the bug.
<img width="2174" height="462" alt="Image" src="https://github.com/user-attachments/assets/3e74d796-967c-447a-ba59-746ebc1a6471" />
I have also experienced this using GPT 5.4 exclusively. Some on fast, but after switching off fast it feels like it was consumed faster.
User ID: user-tq5wuYGXMOUkTgdXx9rH9qhq
Been monitoring this thread since the beginning. I think the issue is pretty widespread. I am also affected. Issue started happening in the last few days.
I do have subagents experimental feature turned on, but I've had it on for more than a month now - and the abnormal usage burn only started recently. I didn't change anything other than switching from 5.2 high as my default to 5.4 high. I use the CLI and don't have fast mode turned on.
I would say the usage is dropping 2-3 times as fast as usual. When I first noticed it, I even started going through 5.4 model docs because it seemed to be using up my quotas way too fast. I can only imagine what it would be like when the 2x promo goes away.
Does everyone need to manually provide their user ID to get a reset?
weekly limit hit in a few hours when i've never hit the daily 5 hour limit using 5.3 medium
user-AdT2JkROSAg7jp824IqPqQpJ
Decided to try out 5.4 xhigh and 5.3 codex xhigh and the rate of credit consumption does seem to be a lot higher than it should be. Not sure if it's really resolved.
user-YjEUQHoYUgHdR67QqpTjxOu6
!Image
This is my first session this week. Extrapolating this numbers after only three 5h sessions I will be on 90% weekly usage. Previous week was completely not like that.
I'm on plus plan using 5.2 medium.
user-86KszM7k18M9DWtXOO064OTK
+1, very fast usage for me.
user id: user-h4JNXBEoAvvY8hXwB8KtZK6u
+1 using 5.4 burned my usage away.
user-IiGyWRofOW5Fmd6nDSovzv73
+1, pro account - burning like never before..
user id: user-NCVWZU6Z9ctgK6cN96xxCLkh
Another, burned on 5.4
UserID: user-Wbc1qMBUVf7T91osN6HLL7rg
Honestly no idea, hard to predict when and how should I use it to maximize the plan. It sometimes felt to drop fast on simple queries, or not move at all with several agents running. Then it like dropped fast and was reset again this week? My first OpenAI subscription, so I had no idea what to expect, but at least it would be highly appreciated to get an e-mail or notification. By luck this topic popped up on reddit for me, I wouldn't know otherwise that there were problems. Thanks!
user-lvKj06VzRe7X3Zwd6StIPPkf
userID: user-qligF2MANrsB2Ggx6vCHv8sv
5.4 burned my usage
Not sure if this is helpful - but 5.4 burned through my weekly usage (and remaining tokens) quickly in 2 days.
user-SvrHMmYCsVIWi0PfDkwE22JR
<img width="1244" height="422" alt="Image" src="https://github.com/user-attachments/assets/c5710752-d646-4337-b138-7711d7f1fb50" />
experiencing the same thing, 5.4 just went thru like 80%+ of usage today, using normally, even tho i switched to "low" halfway through.
id: user-prCBjwoFPi4hmNGc0PNbY0N8
Codex 5.3 has been going pretty smooth for me, but took 5.4 for a spin, in 2 days almost drained my whole weekly limit with my usual usage, whereas the 5.3 codex can have some leftover, appreciate any help.
user-w28KzVw6pjVVVD3i2wR9NyMM
Same issue using 5.4 vs 5.3 Codex using a substantial amount of credits!
user-PqQKQG8v3AJjdZSy0CrZgPAE
Pro Plan and 52% usage in one day. No multi agents running at the same time, just tried fast mode, and turned it off after it consumed 10% of my usage quickly. This can't be right, i take a week to reach 60% of my Claude Max 5x weekly usage. There's something very wrong with weekly limits
user-FjHPVle6X8t8VHQSLxX46p3c
Seeing similar issues. 5.4 did 30% on one prompt and hit my 5 hr limit.
user-c3MwzcPsvBn42EL63daKIexz
Usage has been going down very quickly over the past couple of days. User ID:
user-3HObu82vt03p4xab50sMdDwAUser ID: user-X5ZNosHS0lZnIcyNApR0VpY6 , can confirm 1 planning session on high dropped usage by 20%
weekly limit down to 1% in a few hours on 5.3 medium. have never hit 5-hour limit.
user-AdT2JkROSAg7jp824IqPqQpJ
user-EfwF4BjJUABrRZSVOpTpT6Ya
my limits will stay at a certain percentage lets say 89% then it will stay there for awhile, then suddenly drops to 80%
i used a vpn though, could this be the reason?
Seeing really weird codex usage growth too - somehow managed to chew through 80% of my weekly limit in 2 days :/ part of it seems to be correlated to me updating configs to use gpt-5.4 (high), though it seems way more than '30% higher usage' than gpt-5.3-codex (high).
user-KfUsvRXEeVGDCKzQUxM9mpYd
<img width="705" height="543" alt="Image" src="https://github.com/user-attachments/assets/be9d0320-9d36-4a03-9ab8-2a1fea77bc6f" />
Also noticed a much faster than usual drop in gpt-5.3-codex-spark usage, but don't have hard numbers on token usage per day here.
Pretty worried about this since this is on a pro subscription - after the 2x promo is gone I basically won't be able to use codex if this is genuine.
user-k0UwQaA7TUXXGK3rQIrpvyci
usually no problem with limits, but evaporated my limit on 6., still having 4 days left in the week. just from using gpt-4.
I'm a bit surprised by how quickly my usage is going down. I only made two small changes, did one review, added some documentation, and already used 1% of my weekly limit.
Honestly, it's hard to predict how or when to use it in order to get the most out of the plan. Sometimes the percentage drops quickly even for simple prompts, while other times it does not move at all. This week it suddenly dropped quite fast and then seemed to reset again.
To be honest, that left me a bit caught off guard and somewhat disappointed, mainly because it was not clear to me how this would work in practice.
At the very least, it would be very helpful to receive an email or notification when something like this happens. I only found out about the issue because I happened to see people discussing it on Reddit. Otherwise I would not have known.
user-pxt1wcYoBRnJanlpMUekSa1v
user-FFgFdzcseHa9ApRFO8YNcSah
Also seeing very high usage. I did use some 5.4 but unless the 2x limits have changed from the start this was far faster than +30% usage compared to previous weeks using 5.2 and 5.3 codex
I think I spent roughly 40% of my account usage just yesterday, which is far above what I normally see for comparable work.
From my usage, Codex 5.4 appears to be consuming limits substantially faster than 5.3 under similar workflows. I can’t tell from the client side whether that is due to higher per-turn cost, more hidden intermediate work, retries, or accounting differences, but the observed burn rate feels off enough to warrant investigation.
Account id:
ua-06916063-b4d5-4465-8f1d-f90fa166b9a8
user-qQO0LrL9Yu7KuKvRoR6rXMjg
Captured session identifiers:
<img width="1195" height="323" alt="Image" src="https://github.com/user-attachments/assets/600013fe-bdb1-4291-823e-69c65d2ddd65" />
user-Edh53BgLlgnjWbubYuqBheat
user-WFFujEToGqp78La8P66veY52
Was using GPT 5.2 High before and after the latest update and seemed to have used my weekly limits twice as fast since. Haven't tried GPT 5.4 yet
user-izrKu0zSHA9Kx4rUoGE3jLWU
gone in 24hrs
not a viable product anymore
Same issue for me in codex CLi with 5.4…
<img width="776" height="231" alt="Image" src="https://github.com/user-attachments/assets/2a9894c4-29de-495f-b885-bc34e7e478e9" />
Check this out: Been using 5.3-codex-high and xhigh since Feb all the way up to gpt 5.4's release. the last two bars is gpt 5.4-high and xhigh for around only 14~16 hours of use
its impossible to use gpt 5.4 even with the 2x promo right now, without it I might get like 7~9 hours of use before weekly usage limit is hit!!!!
user-3GKJAL1Ax5UCPt8ZxbKxKjxE
user-2F8o8lSW0ysHCCiG1Kxpf8Sy absolutely melted through 1000 credits and weekly usage even after the double reset cant do more then a day or 2 of coding
Since I'm using Codex, I rarely saw less than 50% of the weekly limits before it resets even when creating large projects and was using 5.3-codex only on High or xhigh. Now I managed to get to 44% weekly limit left in just 1 to 1.5 days with mostly using 5.3-codex high and some requests with 5.4 on Medium for fixing some UI bugs and code optimisation.
user-dA8ZdCNUEaABTXmkUznCh8jp
Used up 60% of my usage with a couple hours today and yesterday.
user-dx9z5U5DtTcIIlQf0ZloKf4d
<img width="719" height="339" alt="Image" src="https://github.com/user-attachments/assets/4f967e08-2b36-4507-b835-52270d2c8782" />
Something is wrong. I haven't changed my context window size and only briefly tested /fast mode, which shouldn't have drained them like this. If this is the normal rate, I'll burn through my weekly limit in a few hours once the promotion ends. This is Pro account by the way.
user-B0qtYBAdzX257eC4Ukyz4B7r
Codex Version: v0.111.0
About 12% of the 5h usage took away 4% of the weekly usage. It would mean that I can only hit 5h limit 3 times per week? That definately seems abnormal... I can understant that 5.4 takes more tokens, than 5.1, but proportion 5h/weekly should be same. Anybody else also observed this?
user-729g2osQwJNK7OPFRY8ZJjdJ
Something is definitely wrong here, the highest day according to the image I barely used Codex compared to the two previous days, yet a large spike in usage.
<img width="346" height="442" alt="Image" src="https://github.com/user-attachments/assets/7cb9ff28-d388-4bad-9e7e-7eff28967b25" />
yes, I can confirm I just did one 5h session fully and spent 32% of weekly usage within it...
User user-nWmJdsZfwinhxMDApQQkizuW , I am below 10% after 3 days and I did not get the reset that was announced. Any help?
Same here.
This shit is eating my tokens like a leon bahhhh
Use user-IdVLIKCGPAmxrfgWKMvpvpc3
I submitted feedback. Usage has been up and down and I am now down to 7% with most of the week left. I missed the benefit of the resets as I took a couple of days off so I actually got penalised (just pushed the reset date out).
2 sessions, 1 yesterday and 1 today, that used 65% of weekly rate. Insane. Used 5.4 and codex 5.3. Default thinking.
i went through 50% in my normal daily usage idk how to get my user id but i was only using codex 5.3 as i have been
I started using my weekly quota yesterday after the reset we had last week, and after 3 sessions I'm exhausted!
Just to fix a minimally designed agent learning system without refactoring.
Considering the 33% increase in usage in GPT 5.4, the consumption was much higher.
Same here: user-IturADd0toT6BKKBm9biTeWo
Same problem here. Went through 64% of my weekly usage in one day, with a Plus account.
Similar issue here on my Pro account. I was using my normal workflows and within a few hours my available usage seemingly tanked.
User ID:
user-2B0QC7ZdudCu1H88xOD5xOpCAbnormal credit usage on Business plan since Feb 27th. The issue has spread from GPT-5.3 Codex to GPT-5.4. My weekly threshold is dropping at nearly the same rate as the 5-hour limit. I've observed a 5x spike in consumption without any change in my actual usage.
id: user-GaUkWAVVXWEz7PBF80JSffhP
Weekly limit literally sank user-e031CoPJ5RzOBLcTqPVl05TG
user-gYHBAvrzgDmS6W1MyiamshBM
Both GPT-5.3-Codex and GPT-5.4 were burning through my usage, I had one prompt that used about 17% last night of my weekly limit in Plus.
Went through 60% of my usage in a couple of hours thursday, using Codex 5.3 medium.
user-NM5vbCjNw1NJy5buCQtm09uX
Yes likewise over the past few days. Really disappointing as I was loving using codex 5.4 I havent even used it on Xhigh yet!
Having the same issue here - my reset limit day jumped from March 9 to 11. I'm assuming this is because my weekly usage jumped incorrectly, too.
User ID: user-FYQ4quM70OGZ4ryRGcjCDI9n
66% usage ate with ~5 prompts
Please review the usage, there is definitly a problem with the USAGE. never had any issues with the usage suddunly we do. also the weekly usage us draining like10x more than the 5 hours. strange !
I completely agree - it feels like the weekly limit has been reduced fivefold.
Just a note that one prompt + plan generation seemed to wipe a huge chunk of my rate limit. Session ID: 019cbfe2-fb85-7013-9f80-361084a03976
UserID: user-MPTBLC31Dx8eNbuRXFwWILXt
I'm burning through my account in roughly half a week now. I used to be quite comfortably within limits before.
Even switched from xhigh to high to no avail.
How do I get refund for my 120 EUR worth of credits this consumed on top of my regular quota? Your Help entry point in Codex just brings me to a chat bot.
user-DvRLyz31uPfJyx95yQtQhKf2
crazy usage today
user-MgvQR9VFolsLteA4R4NuTgzs (whole weekly quota exhausted in under 2 days, never seen anything like it)
Hello!
user-vCJq0VW2uSEmtAVWM6wTz0Hf here, was surprised to see the usage limit alert today (for the first time ever). Appreciate any assistance you can provide.
user-W5Q6Ci7yBt14rgqeSj6JkzNa and user-b6F3W1QHIhyIBhzOKhG1D86q
Weekly usage drop very fast. More than 20% usage of the weekly quota on both account with 5.3 codex for barely 1 hour of coding
Something feels very off
user-kJrkqizLtJdunfXb8fyT12cq
This is really unsustainable. I can’t use it the same way as before — in almost 2 days I used what I used to use in 1 week. I honestly can’t believe it.
pro plan noticed I drained approximately 4x faster despite using less intensive tasks
user-7W2aLIiOqCzamw7dNUbs9mH6
same here as all of the above.
context
Same here, 75% of weekly usage in just 2 days.
user-34JR1FkGGJamgmKe4pgMUoG2
user-W9hxpZI4Dp85DQnvDl8hzg4p
i'm doing per usual tasks as prior on 5.2, 5.3 both on Extra High never ever reached weekly limit on heavy use. and a max of 2 new threads a day.
migrated to 5.3-codex Extra High
i have reached all 3 token-context windows on a pro plan in a single day, and reached my weekly limit already 1 and half days going into the new week
Both my accounts have been. It’s insane but particularly this one almost 50% in like 2 hours which would last me half a week. user-aTaujlF3KWKDePlACvmTsqkw
And user-ydA9yAl3rJqBCNGmuVh82Orw
feels like this incident should be re-opened https://status.openai.com/incidents/01KK26XE1W536H7DQV2EXM3GHE?
user-kTtyDmbwLLBtRyLocjte5Nd4
Also experiencing the same issue!
The weekly usage is the crux of the issue for both Pro an Plus users, you guys really need to check this
same here this is nuts
Still not fixed for 2 hours it drained my weekly to 50% that is nuts, also the 5 hours limit is literally melting fast.
I turned off websockets and switched to js_repl only and I've noticed it helps a little bit compared to when I had websockets enabled, but it's still just DRAINING. 1 singular prompt took away 1%, and the prompt was simple and didn't require new reading.
Watched my weekly usage drop from 100% last night to 50%, then from 50% today to 20%, meanwhile in neither session did I reach the 5 hour limit of the session.
I’m seeing this too, quota is depleting unusually fast during light use.
User ID: user-DBL1mFXGX9BVFpfX7OioQ4rj
It's pretty clear that this is affecting many users. We just reset rate limits for all Plus and Pro users, so you should be back to 100%. Apologies to everyone who has been affected by this.
There's no need to post more user IDs or bug reports. We're very aware of the issue, and we're working over the weekend to try to get a fix in place as quickly as possible.
Thanks, and what about the team/business users?
@etraut-openai Thank you!
@etraut-openai thank you, much appreciated! 🙏
@etraut-openai thanks, but team/business users have not been reset
@etraut-openai Thank you so much, you really saved the day.
Thank you
@etraut-openai Please don't forget the team/business users, we haven't even received the first reset from days ago, and now we haven't either
maybe its something that your company needs to deal with, its an enterprise plan not personal, so your company need to open an issue
It doesn't seem the most practical for each company to have to open an issue when limits are being restored on a massive scale, but thanks for the input
companies are the ones that pays for those plans, not you, so its obviously something that you need to get scalated
I understand, thanks for letting us know what we need to do and that we have to wait until Monday to start the limit request process.
Please hurry up with the 100 dollar plan so that I can switch from Claude, thank you guys you are the best
@etraut-openai Thanks a lot! A final summary about the cause would be nice!
Trend from yesterday:
<img width="2496" height="828" alt="Image" src="https://github.com/user-attachments/assets/33dc24f9-4360-4300-aa3b-be284a6fbae6" />
The smallest bars are my non-working days. Just still monitoring if the usage is higher and whether GPT4 justifies the 30% bump in usage.
I've been using sub-agents since the beginning. Planning on
5.4. xhigh, multiple iterations and explorations with sub-agents. Then switching to worker sub-agents5.3. codex highto implement.Yesterday, overall +30% usage but it has been growing steadily. Just worried if this is a new normal with 2x limits, even PRO plan doesn't seem like enough considering I was barely hitting weeklies at all.
UserID:
user-EwwvXRbDXkpKEnwW9TeVbKGTQuick note, @etraut-openai : my Codex usage reset about 5–6 hours ago (around two minutes after your last message), and I’m already back at 93% of weekly usage (PRO plan). That is far faster than what I was seeing before the reset ever.
What makes this concerning is that I’m only iteratively debugging a single project in one thread, using standard context length. Something seems wrong. On 5.3 xhigh, reaching this same usage level usually took running three projects in parallel for roughly 12 hours straight. My rough estimate is that 5.4 is consuming closer to 6x the usage, not 2x.
I could personally accept 2x to 2.5x higher usage on 5.4 xhigh versus 5.3 xhigh, since the model is better and faster, and there is real infrastructure cost for you. I could work around my workflow complexity in deciding when to use standard, when to use fast, and when to rely on other models. But at around 6x, Codex starts to lose its main competitive advantage: affordability to run top tier models.
I use windows app.
My config is:
model = "gpt-5.4"
model_reasoning_effort = "xhigh"
personality = "pragmatic"
[windows]
sandbox = "elevated"
My app settings:
Default open destination:
File Explorer
Agent environment:
Windows native
Integrated terminal shell:
PowerShell
Language:
Auto Detect
Thread detail:
Steps with code output
Require ^ + enter to send long prompts: (untoggled)
Speed:
Fast
Follow-up behavior:
Queue
I haven't changed any defaults from config or app settings apart from agreeing to fast when pop up offered it. And I don't use and don't have any MCP servers connected.
I'd like to work on this issue. I've analyzed the codebase and have a fix ready.
@etraut-openai This is unbelievable at this point. For the third time in a row we haven't received the reset either. I'm administrator in my organization, went to their customer service chatbot and got replied something like a Customer Support Specialist would get in touch with me. I can't understand why they haven't reset for everyone, especially when this is a bug affecting all of us.
This is after usage reset. Every thing started from 100% on both 5 hour window and weekly. my usage drained in the background without any tasks. my last prompt output and I still had 93% weekly, 60% on the 5 hour. nothing was running in the background, no other sessions active. I came back to my session and I have 0 on the hourly and 70% on the weekly
!Image
!Image
Session: 019cbfd9-3a1a-7153-97cf-dc2680414f7e
I'd like to work on this issue. I've analyzed the codebase and have a fix ready.
<img width="2566" height="728" alt="Image" src="https://github.com/user-attachments/assets/f4286e66-21a8-4e2f-a1ed-c112a98661d0" />
I audited my local
~/.codexdata (state_5.sqlite+ rollout JSONL) to understand my own March 5-6 spike, and I rebucketed it intoAmerica/Los_Angelesso it aligns more closely with the online usage graph.This is just my personal analysis of my personal spike. I’m not claiming it explains the whole GitHub issue or any other reporting/accounting problems other users may still be seeing.
What I found:
spawn_agentcallsspawn_agentcallsThis does not look like “I just used subagents a lot.” Earlier bounded counterexamples already had real multi-agent usage:
0.104.x0.106.0and0.107.00.107.0The strongest comparison is one long-lived root thread:
019cac4b-28ac-7280-82af-613e60ec2e9c38,211spawns, about23.5Btokens31,077spawns, about23.2BtokensSo from my local data, this looks more like runaway fan-out in a small number of long-lived multi-agent threads under newer CLI/runtime behavior, not just “subagents were enabled.”
The version progression is what makes this stand out to me:
0.104.x,0.106.0, and0.107.00.110.00.111.0Based on the release notes and code, my current hypothesis is that
0.110.0is the important behavioral shift: it appears to have both encouraged more aggressive parallel subagent delegation and reduced subagent startup/reliability friction.0.111.0may have amplified throughput, but it does not look like the main semantic change.I also checked whether switching threads from
gpt-5.3-codextogpt-5.4was the trigger. It does not look like it:gpt-5.3-codex -> gpt-5.4accounted for 70,419 spawnsgpt-5.4turnSo my current read is that this does not look like a simple “5.4 is more expensive” explanation, and it also does not look like
fork_contextalone explains it.If useful, I can extract more data and share the specific audit script I used. I’d prefer not to post exact session transcripts or other raw per-session artifacts publicly in the issue thread.
@etraut-openai Why is it that Business/Team members are not reset like Plus/Pro members? I pay if not more, then the same as a Plus user and is affected by the usage bug in the same way. It shouldn't matter what plan you are on, when everyone is affected, just because one part is primarily private and the other is corporate paid for.
Hey what's going on with usage? It reset, I was down to 70% (again), it reset (again), and now I'm down to 68% AGAIN.
Ain't no way I used 62% in a day. I've done that before and it involved 9, 10 simultaneous sessions. I was using 2 for this. No wild amounts of sub agents, etc.
Looking at token counts in session logs I'm seeing absurdly large values repeated (as in the exact same value) duplicated across multiple files 6, 8 times. That 6-8x multiple sure seems to match up with the amount of usage that is getting nuked.
This feels more like 0.2x weekly limits than 2x weekly limits.
<img width="1199" height="292" alt="Image" src="https://github.com/user-attachments/assets/33e8d00c-fa36-4183-be18-400238483331" />
---
If I had to make an educated guess, since you have already zeroed in on websockets being one potential cause, and I had seen multiple "Reconnecting..." messages, with a number of reconnection attempts that loosely correlate with the duplicate token counts: However you are persisting state from one websocket connection to the next is not correctly invalidating the states of previous connections on either? end
Been doing networking bullshit a long time, since back before websockets existed. I've seen this exact same thing happen in isp/vpn billing, automotive telemetry, and video encoding/routing contexts.
After scrolling up through 300 comments I'm seeing you instruct us to not use 5.4 because it's muddying the waters?? Buddy, if your logging and metrics are not up to the task that's on you. How about you stop muddying up the waters by distracting from the issue at hand with dismissive non-causes. It should be immediately obvious to you and anyone else who can do basic math that the increased usage of 5.4, which is a known value (inferrable from API pricing), is only a fraction of the multiplier we are seeing usage drain.
---
Why are emploidcels shitting up a real issue? Make your employer handle it through the correct channel tia
@etraut-openai Why are we Team/Business users receiving different treatment and being ignored on the two occasions that the limits have been restored? It is impossible that with the countless times this has been complained about, you haven't seen it. You are ignoring us for some reason that I don't understand
Why are people surprised it's still draining fast? They already acknowledged the issue, they just did a reset so people can keep using but they haven't fixed the issue as they're still investigating, so the drainage will continue until then... it's not like the reset would fix the drainage, it's just a remedy
Well at least they reset the quota to investigate, unlike Anthropic. Thank you OpenAI and Codex team!
Not trying to shit up a real issue, just plain wondering about the difference in treatment. Im the owner of the Team and I have raised the question through support, but I dont expect im getting a straight answer as to why the two groups are treated different. The problem is not affecting me any less than everybody else, just beacuse im a business member.
I'd like to work on this issue. I've analyzed the codebase and have a fix ready.
I computed my limits (on Plus plan) to be 75M for 5H and 250M for weekly. Anyone can share their results or even know what the limits were before the x2 promo?
Honestly by looking at my weekly tokens/estimated cost since September, I don't even think I have the x2 promo applied. It matches my experience, I never felt like my limits become better. Initially I upgraded from Medium to High but I immediately noticed limits going down faster so I reverted to Medium. I believe I've been using Codex more often but I'm also hitting limits which rarely happened before and I don't think I'm using 2x more. I even mentioned this on Reddit on February 4th but most people didn't feel the same experience so I just thought it was my usage... but apparently not.
<img width="1012" height="463" alt="Image" src="https://github.com/user-attachments/assets/9aa08fa5-3cac-4985-8b28-25c5a16c9d2a" />
<img width="1002" height="455" alt="Image" src="https://github.com/user-attachments/assets/755ba134-10f6-409d-a6ff-1cbd79045c5d" />
After the reset, I just drained 50% of weekly usage in just two 5h windows, with got 5.3 codex medium and 2x usage limits in codex app... this still feels off
Thanks for all the bug reports. We are aware of this issue and are working on a fix.
We don't need additional reports at this time. I'm going to lock this issue temporarily.
Thanks for your patience as we worked through this issue.
As I mentioned above, we identified a problem that explains some of the behaviors described in this thread. It resulted in erratic usage reporting. Some users reported that the remaining usage sometimes appeared to drop suddenly or oscillate between a higher and lower value. Other users reported inconsistent usage stats between the local client and the web usage dashboard. These problems were due to a syncing issue between regions (data centers). The problem disproportionately affected users who are geographically located on boundaries between regions.
We've added a mechanism to reduce sync latency between data centers. If your requests happen to be routed to different regions, you may still see minor differences in usage stats, but it should never be more than a few percent and only over short time windows.
We don't think that this issue explains all of the reports of higher-than-expected usage. We investigated many such reports and didn't identify any other problems with usage accounting. We think that these are largely explained by a combination of the following:
gpt-5.4, this will consume usage about 30% faster thangpt-5.3-codex.gpt-5.4), this will consume usage at a 2x rate.multi_agentfeature, subagents will typically consume tokens at a higher rate than if you are not using subagents. As we continue to refine this experimental feature, some changes may affect token consumption. If you are concerned about usage consumption, leave this feature in its default (disabled) state for now.If you think that you are seeing higher-than-expected usage consumption that is not explained by the above and have strong evidence and/or analysis to back it up, please let us know.
I'm going to close this issue and keep it locked because it's already so long that it's hard for Github to render.