Weekly limit is draining like the old 5-hour limit
What issue are you seeing?
Since the 5-hour limit disappeared, the weekly limit is dropping at about the same speed as the old 5-hour limit did.
I am using the same normal setup as before, GPT-5.5 High. I am not using Sol or any unusual heavy mode.
Before this change, a normal work session mainly reduced the 5-hour limit and the weekly limit lasted much longer. Now the weekly allowance drops very quickly during the same kind of work.
What steps can reproduce the bug?
- Open Codex in VS Code.
- Select GPT-5.5 High.
- Use it normally on an existing project.
- Check the weekly usage before and after the session.
The weekly percentage drops at roughly the rate the old 5-hour limit used to drop.
The same weekly usage value is also visible in ChatGPT Settings, so this does not seem to be only an extension display bug.
What is the expected behavior?
Removing the 5-hour limit should not reduce the total weekly capacity.
The weekly limit should still last roughly as long as it did before, instead of behaving like a single old 5-hour allowance.
Additional information
I first noticed this after the 5-hour limit disappeared from the usage page in July 2026.
My model choice and normal workflow did not change. This happens with GPT-5.5 High, so Sol is not required to see the problem.
Please check whether the weekly allowance was reduced or if the usage accounting is wrong.
16 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
You might have too many skills or MCPs. Delete those "programming-language-skills" and any obvious thing that GPT already had in it's inference. Most skills in marketplaces are not useful for frontier models, but might be useful for models with less deep reasoning.
Or your computer might be compromised, there are report of malware that steals AI token usage by proxing botnet requests through your subscription
The previous two comments are unrelated to the reported issue and are unsupported speculation. I noticed the weekly limit started draining at roughly the same rate as the former 5-hour limit when the 5-hour limit disappeared. Please keep the discussion focused on the reported usage-accounting problem and verifiable evidence.
I commented that because I cannot reproduce this bug here.
I am experiencing the same issue in the Codex desktop app.
Account and environment
What happened
My weekly usage allowance reset on the morning of July 16. After the reset, I used GPT-5.6 Sol for a task that ran for approximately two hours. When I returned in the evening, I found that only 18% of my weekly allowance remained.
This was extremely unexpected. Under my normal usage pattern, a comparable two-hour task would usually consume less than 5% of the weekly allowance, and often no more than about 2%.
I took screenshots of the usage level and then opened a short discussion using GPT-5.6 Terra in High reasoning mode to investigate the issue. During that discussion, the remaining allowance continued to decrease by roughly 1 percentage point after every one or two messages.
The session was not unusually large or long. The total context was only about 45,000 tokens, and the model spent roughly 30 seconds reasoning. There had only been a brief discussion of a folder before this. Therefore, the rate of depletion is difficult to explain as a normal consequence of a long context window or an unusually expensive task.
After only a few additional exchanges, my remaining weekly allowance fell from 18% to 10%.
Expected behavior
The weekly allowance should decrease at approximately the same rate as it did before this change. A two-hour task and several short follow-up messages should not consume around 90% of a USD 200/month Pro account's weekly allowance.
This appears consistent with the report in this issue: after the five-hour limit disappeared, the weekly allowance began draining at a rate comparable to the former five-hour allowance. Please investigate whether the weekly quota size or the usage-accounting logic changed on July 16, or whether usage is being counted incorrectly.
After the short GPT-5.6 Terra High conversations described above, the Codex usage page still showed 10% remaining. I left the page open for more than ten minutes and the value remained unchanged at 10%. There were no visible active or queued Codex tasks.
I then closed my laptop and travelled home. During this period, I did not use Codex, ChatGPT Work, ChatGPT for Excel, or any Workspace Agent. I only used a normal ChatGPT conversation, in which ChatGPT made one GitHub connector action to post a comment to this issue.
When I later reopened Codex, the remaining allowance had fallen from 10% to 0%, and the application displayed “Your Codex and Work usage allowance is exhausted,” with the next reset shown as July 23 at 05:16.
A normal ChatGPT conversation and a GitHub connector action should not, based on the published documentation, consume the Codex/Work shared agentic allowance.
Agreed wholeheartedly. If this is a transparency regression, then add back in the 5-hour timers (👀 Tibo).
Nobody likes being peed on their leg and then being told it's actually Apple Cider when it's really not.
All the "oops I did it again" goodwill reset vibes should mathematically square with the most recent capacity equivalents of at least EIGHT full 5-hour windows per week when translated for the suspended sub-timer.
Token-use acceleration, hidden behind false largesse, will be the fastest way to lose Pro customers (👀 Tibo).
Update: this issue is still ongoing. Since the 5-hour limit disappeared in July 2026, my weekly allowance has continued to drain at an unrealistically fast rate with my usual GPT-5.5 High workflow. My model, settings and normal usage pattern have not changed.
If it helps anyone quantify what's happening: the information is already on your machine.
The rollout files under
~/.codex/sessionsincludeused_percentfor each quota window (5-hour, weekly, monthly), along with their reset times. I surface those in unsnooze usage, which renders each window as a progress bar with its exact reset date, so you can actually watch the weekly quota drain instead of only discovering you've hit the limit.It's been much nicer than guessing how much budget is left.
https://github.com/saaranshM/unsnooze
I agree with this and this is very obvious even without clear numbers. It is such a massive difference to before that you merely breathe in and lost 10% already, whereas when I resubbed July 7th, I had been working normally on a 150k line code project for easily 14-18 hours straight and I did not even go below 90%.
I would not say the new and current state it is roughly the 5 Hour limit but maybe like 50% slower than that. But maybe I am mistaken, didn't check yet since I was too busy.
Same workload, etc. as when I first resubbed to Codex July 7th or so, but weekly limit by now drains much faster with the same workload.
However I am getting random resets in the middle of the "week" and my reset date keeps shifting backward, so I am not that bothered yet. Even though this particular time I am quite literally having my weekly limit *exhausted 1 day into the "week" given the date kept shifting with the free irregular resets.
Now given the frequency of those, I assume I will get another random reset for no apparent reason before that actual weekly reset date.
But we will see. If not, then this would obviously be horrible given nothing in terms of the workload changed.
<img width="1920" height="415" alt="Image" src="https://github.com/user-attachments/assets/a5aa8b75-e74a-405f-9a3a-5b032c29ba1d" />
Turns out it increased a little bit on average, since I had two workflows running for two of those days but compare this with early days of july, as I said, I barely got even below 90% weekly usage in the first 2-3 days combined.
Then after those "free resets" in the middle of the week, 2 days ago or so I ended up using 100% of my weekly allowance in just one day. lol
To me that is not explicable by this graph.
Title: Weekly usage quota depleting significantly faster since the 5-hour limit was removed (July 2026)
Description:
Since the temporary removal of the rolling 5-hour usage limit on July 12, 2026, my weekly usage quota is draining at a much higher rate than before, despite no change in my usage pattern.
What I observed:
Expected behavior:
Removing the 5-hour limit should not reduce total weekly capacity. The weekly allowance should last roughly as long as it did before, not behave like a single old 5-hour bucket stretched across a week.
This appears to match a known issue already reported by other users (see openai/codex issue #33685), where the same behavior was observed independent of using Sol or any unusual workload.
Title: Weekly usage quota depleting significantly faster since the 5-hour limit was removed (July 2026)
Description:
Since the removal of the 5-hour usage limit on July 12, 2026, my weekly usage quota is draining at a much higher rate than before. Today it 'broke the record'.
What I observed:
The daily/5-hour usage indicator has disappeared from the usage page.
The weekly quota is depleting faster than it did when the 5-hour and weekly limits operated separately.
Today I've depleted the limit that was going down 'steady' but faster than it should. I've clicked one of the free limit reset, after running 2 tasks that took aprox 1h in total I've somehow drained 89% of this reseted limit, this is unacceptable when I'm paying for PRO subscribtion. Since release of 5.6 models I have issues - similar tasks take HOURS instead of minutes, response time for simple questions takes minutes instead of seconds, and how this quota.
Expected behavior:
Removing the 5-hour limit should not reduce total weekly capacity. Resetting limit should bring back full weekly allowance. The weekly allowance should last roughly as long as it did before, not behave like a half of old 5-hour bucket stretched across a week.
Can confirm that something fundamentally is off with the usage. Here is my weekly burn down chart for today. The last days were already fast but today it burns my usage like crazy with not even complicated or long running tasks. It crashes down like it forgot to put on its parachute
<img width="479" height="325" alt="Image" src="https://github.com/user-attachments/assets/d5ce88c8-eb5d-4778-89c4-70844fd56b64" />
Adding another detailed data point from Codex Desktop.
Environment
com.openai.codex26.715.52143, build5591First event: nearly an entire weekly allowance consumed in one day
After the GPT-5.6 rollout, I intentionally used GPT-5.6 Sol High and Very High during one intensive working day.
The work consisted of several substantial real-world agentic workflows involving:
This was not a large number of trivial prompts. It was a relatively small number of prompts that each launched significant autonomous work.
By the end of that single day, almost the entire weekly allowance had been consumed.
GPT-5.6 Sol produced noticeably better results than GPT-5.5 and completed more useful work, so some increase in consumption is understandable. However, consuming almost a complete weekly allowance during one intensive day makes sustained professional use difficult.
GPT-5.6 Terra Very High also appeared to consume the weekly allowance very quickly, despite Terra being positioned as the more economical GPT-5.6 option.
Second event: very high consumption while using GPT-5.5
The allowance subsequently returned to 100%, with the next reset shown as July 26, 2026.
During the following working day, I deliberately returned to GPT-5.5 Very High rather than continuing to use GPT-5.6 Sol.
Even while working primarily with GPT-5.5 Very High, almost 70% of the newly reset weekly allowance was consumed during that day.
Early in that period, the displayed allowance had already fallen from 100% to 78% after a relatively small number of prompts, some of which launched substantial agentic workflows.
I checked the meter again more than five hours later and again the following morning. The consumed usage did not recover, and the reset date remained July 26.
This confirms that the visible meter was showing the weekly allowance rather than a temporarily mislabeled 5-hour window.
The important point is that this unusually high consumption also occurred with GPT-5.5 Very High. Before the GPT-5.6 rollout and the recent changes to Codex usage limits, comparable GPT-5.5 Very High work did not consume anything close to this proportion of the weekly allowance.
Why this looks like more than normal GPT-5.6 model cost
I understand that one prompt may trigger many model turns, tool calls, context reads, reasoning steps, retries, verification operations, and compactions.
The concern is therefore not the raw number of prompts.
The concern is the observed change in effective weekly capacity:
This makes it difficult to determine whether the cause is:
Expected behavior
Questions
I am intentionally not including account identifiers, session IDs, private repository details, infrastructure information, production logs, or screenshots containing sensitive data.
This is only some speculation about the problem, and I allow myself to post it here because I opened this issue.
I think the timing of the GPT-5.6 rollout makes this quota problem much harder to notice and understand.
Sol and Terra were introduced as more expensive models, so when the weekly usage suddenly started draining faster, it was easy to assume that the new models were the reason. There were also many reports about unexpectedly high Sol consumption, so this explanation became even easier to accept.
But this mixes two different things together: the normal higher cost of a new model, and a possible change or bug in the weekly quota itself.
This is why reports from users who stayed with GPT-5.5 are especially important. If the same GPT-5.5 workflow now drains much more weekly allowance than before, then the price of Sol cannot explain the whole problem.
At almost the same time, the 5-hour meter disappeared. This also removed the easiest reference point users had for seeing how quickly their usage was being consumed.
So now we have new and more expensive models, the removal of the 5-hour meter, and much faster weekly depletion happening close together. This makes it very difficult to separate expected model cost from an accounting problem or an undocumented reduction of the weekly capacity.
I am not saying this was necessarily intentional. But the timing and the lack of detailed usage accounting make a serious quota regression very easy to hide behind the expected higher cost of the new models.