Repository navigation
[Bug] ChatGPT Plus quota consumption increased ~9x between comparable Codex App tasks #41957
Description
What version of the Codex App are you using (From “About Codex” dialog)?
26.825.51511
What subscription do you have?
ChatGPT Plus
What platform is your computer?
Microsoft Windows NT 10.0.22631.0 x64
What issue are you seeing?
ChatGPT Plus quota consumption in Codex App increased dramatically between two comparable consecutive tasks.
On 2026-08-31, a ~24-minute design task consumed 8 percentage points of my 5-hour quota.
On 2026-09-01, a very similar ~53-minute design task consumed 75 percentage points of my 5-hour quota.
Local Codex session telemetry shows that input token usage increased only from 13.56M to 22.04M (~1.63x), while 5-hour quota consumption increased from 8% to 75% (~9.4x).
Both tasks were part of the same design workflow, used 4 active threads, and had very high cached-input ratios.
The second task also increased my weekly quota usage from 21% to 32% in a single run.
What steps can reproduce the bug?
Environment
- Codex App: 26.825.51511
- ChatGPT subscription: Plus
- OS: Windows x64
- Local Codex CLI/runtime: 0.151.0
- Model observed in local Codex session JSONL: gpt-5.6-sol
Reproduction
I was working on consecutive sections of the same software design document in the same Codex workflow.
Task 1 — Section 5.10
Prompt:
接着做5.10的设计
English meaning: "Continue the design of section 5.10."
Time:
2026-08-31 19:08:43 to 19:31:56 local time
Duration:
~23m57s
Local session telemetry:
- Active threads: 4
- token_count updates: 97
- Input tokens: 13,555,743
- Cached input tokens: 13,231,104
- Non-cached input tokens: 324,639
- Output tokens: 43,307
- Total tokens: 13,599,050
- Cached input ratio: 97.61%
Quota change:
- 5-hour usage: 58% -> 66%
- 5-hour delta: +8 percentage points
- Weekly usage: 19% -> 21%
- Weekly delta: +2 percentage points
Task 2 — Section 5.11
Prompt:
继续5.11的设计
English meaning: "Continue the design of section 5.11."
Time:
2026-09-01 08:21:37 to 09:13:10 local time
Duration:
~53m26s
Local session telemetry:
- Active threads: 4
- token_count updates: 194
- Input tokens: 22,037,379
- Cached input tokens: 20,967,936
- Non-cached input tokens: 1,069,443
- Output tokens: 89,572
- Total tokens: 22,126,951
- Cached input ratio: 95.15%
Quota change:
- 5-hour usage: 0% -> 75%
- 5-hour delta: +75 percentage points
- Weekly usage: 21% -> 32%
- Weekly delta: +11 percentage points
Comparison
Input tokens increased only:
13.56M -> 22.04M
This is approximately a 1.63x increase.
However, 5-hour quota consumption increased:
8 percentage points -> 75 percentage points
This is approximately a 9.4x increase in quota consumption.
Very roughly normalized by locally recorded input tokens:
- Task 5.10: ~0.59 quota percentage points per 1M input tokens
- Task 5.11: ~3.40 quota percentage points per 1M input tokens
That is approximately a 5.8x difference in quota consumption per locally recorded input token.
Weekly quota also increased disproportionately:
- Task 5.10: +2 percentage points
- Task 5.11: +11 percentage points
The weekly usage is continuous across the two tests:
Task 5.10 ended at 21%, and Task 5.11 started at 21%.
For Task 5.11, the first recorded 5-hour usage was 0%, and it reached 75% during this single task.
I understand that local token counts may not map linearly to ChatGPT subscription quota. However, this difference appears much larger than the difference in workload and is reproducible in the local session telemetry.
Session ID:
01a017fb-2df4-76e0-a295-cfb279dfb511
At the end of the affected task, /status showed approximately:
- 5-hour quota remaining: 25%
- Weekly quota remaining: 68%
- Context remaining: 67% (~85,896 / 258K used)
What is the expected behavior?
Quota consumption should remain reasonably consistent for comparable workloads using the same subscription and model.
I do not expect quota usage to map linearly to locally recorded token counts. However, an approximately 1.63x increase in input-token workload should not result in an approximately 9.4x increase in 5-hour quota consumption without some corresponding change in workload, model, or documented quota behavior.
I would expect the second task to consume more quota because it ran longer, but not 75% of the entire 5-hour quota in a single comparable design task.
Additional information
No response
Activity
github-actions commentedon Sep 1, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
t-var-s commentedon Sep 2, 2026
I’m seeing a very similar discontinuity on ChatGPT Business / Codex Desktop / macOS Intel, also using GPT-5.6 Sol.
My detailed reproduction is in #20153, but the sequence after my 5-hour quota reset was:
Fast mode was disabled.
What seems relevant to this issue is the discontinuity: the complex existing thread plus compaction used 7 percentage points total, while the fresh-thread task after an app restart coincided with an 84-point drop.
I don’t have the same local token telemetry comparison as this report, so I can’t establish the charging ratio directly, but the user-visible behavior looks very similar and occurs on a different subscription tier and OS/platform.
Full details: #20153
LimpingNinja commentedon Sep 3, 2026
Adding support for this, I took this sample set from another direction to provide additional support and detail, here's an 11% chunk of Sol Ultra (last week) vs an 11% chunk of Sol Max and Terra Xhigh; the latter (lesser models) provided significantly less wall time and value.
Exact same-band comparison
Window Remaining usage Recorded model / effort Duration Local token-usage updates Raw tokens Model rate-card proxy*
Aug 26 64% → 53% GPT-5.6 Sol / Ultra only 1h 16m 604 229.4M 2,753.3
Sep 2 64% → 53% GPT-5.6 Sol Max/inherited, plus 5h 33m 885 91.3M 1,316.3
33 Terra xhigh updates
For the same 11 percentage points of quota consumption, the Sep 2 interval had:
60% less raw token work: 91.3M vs 229.4M
52% less token-weighted work under the published model rate card: 1,316.3 vs 2,753.3
47% more local token-usage updates: 885 vs 604
73% fewer tokens per update: approximately 103K vs 380K
The Sep 2 workload consisted of many smaller agent/tool-loop requests, not more token-weighted work. The local records show 885 token-usage updates during the interval. #42448
Adding to multiple tickets to spiderweb them.
razdgann commentedon Sep 3, 2026
This is exactly the kind of discrepancy that makes request-level observability important.
If token volume increased ~1.6x while quota consumption increased ~9x, aggregate session totals aren’t enough to explain what actually happened.
I’d want to compare each request by cached input, uncached input, output, context size, model, duration and estimated cost/quota impact, then identify the specific requests where the behavior diverged.
We’re working on this from the gateway side with ModelGate. One of the goals is to make cost anomalies attributable to individual requests instead of discovering them only after the session or bill looks wrong.
The telemetry you’ve included here would make a very useful test case.
Bananan4eg commentedon Sep 6, 2026
I’m seeing the same issue on ChatGPT Plus / Codex Desktop for Windows.
The 5-hour quota started draining dramatically faster than before, even on relatively small tasks.
My measurements today:
This is very different from my previous usage. Until very recently I could spend much longer on larger Codex tasks, including building and testing an entire project, without exhausting the 5-hour window.
The problem also persists after isolating the workspace, so it does not appear to be caused only by the size of the project/context.
This looks like a possible regression in quota metering/accounting or a significant unannounced change in how usage is calculated.
Platform: Windows
Subscription: ChatGPT Plus
Model: GPT-5.6 Sol
Reasoning: High
I did not capture timestamped before/after screenshots for every run, but I manually recorded the quota percentages while testing.
EaglePPP commentedon Sep 6, 2026
Did all of you use codex offical apps and addons? I am not sure if this is relevant, but about 22-24hrs ago I switched from Zed's Codex ACP to Zed agent and tried Astra there(not sure if there is any correlation) then the usage started to burn like crazy. I would say ballpark feels like 10x faster usage even on luna and sol medium.