Weekly quota dropped from 21% remaining to 0% while idle; reset timestamp changed unexpectedly

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Windows 11

What issue are you seeing?

I am reporting what appears to be a Codex weekly usage accounting / reset reconciliation inconsistency. I have local rollout telemetry with exact server-reported quota snapshots before the failure.

The strongest evidence is the following timeline (CEST / UTC+2):

• 2026-09-09 14:52:45 — the Codex rollout reported approximately:

  • 5-hour limit: 100% used
  • weekly limit: 79% used (~21% still remaining)
  • 5-hour reset timestamp: 2026-09-09 19:29:51 CEST
  • weekly reset timestamp: 2026-09-15 08:17:53 CEST

• Between 14:52:45 and 19:08, this rollout contains no further agent work or token generation.

• At approximately 19:08, the Codex UI showed the 5-hour allowance at 100% remaining, even though the earlier server-reported reset timestamp was still about 22 minutes away.

• At the same time, the weekly allowance had suddenly become 0% remaining.

• 2026-09-09 19:08:15 — the next request was rejected almost immediately with usage_limit_exceeded, before any meaningful model work occurred.

• The retry message now said to try again on Sep 12, 2026 at 4:17 PM, despite the earlier server telemetry saying the weekly reset was Sep 15, 2026 at 08:17:53 CEST.

So two unexpected changes appear to have happened without a corresponding foreground task in this rollout:

1. Roughly 21 percentage points of weekly allowance disappeared.
2. The weekly reset date changed from Sep 15 to Sep 12.

There is also an apparent short-window inconsistency: the 5-hour allowance appeared restored before the previously reported reset time.

I cannot determine from local telemetry whether this was caused by delayed server-side reconciliation, incorrect attribution from another surface, a meter/display synchronization problem, or an entitlement/rate-limit bucket transition. If another Codex/Work/background surface consumed the missing allowance, please identify the server-side activity responsible so the usage can be audited.

This looks related to the broader abnormal quota/accounting reports tracked in #41220, and is similar in symptom shape to recent reports such as #44205 and #44206.

I can provide the affected rollout JSONL privately/sanitized if useful. It contains the exact token_count / rate-limit snapshots around the transition.

What steps can reproduce the bug?

I cannot intentionally reproduce the accounting jump, but the observed sequence was:

1. Use Codex normally until the 5-hour bucket is exhausted while the weekly bucket still reports ~21% remaining.
2. Stop foreground work in the affected rollout.
3. Wait several hours.
4. Re-open/check Codex usage before the previously reported 5-hour reset time.
5. Observe the 5-hour allowance restored while the weekly allowance has dropped to 0%.
6. Submit a new request.
7. The request is rejected immediately with usage_limit_exceeded, and the reported weekly reset date differs from the earlier server-reported weekly reset timestamp.

What is the expected behavior?

• Remaining weekly allowance should not disappear without attributable usage.
• A weekly reset timestamp should not silently move from Sep 15 to Sep 12 unless there is a documented bucket/entitlement transition.
• The 5-hour allowance should reset according to the server-reported reset timestamp, or the UI/server should explain why it changed.
• Usage should be auditable enough to identify which foreground/background task consumed each quota bucket.

Requested resolution / compensation

Please inspect the server-side usage ledger and rate-limit bucket history for this account around 2026-09-09 14:52–19:08 CEST and determine exactly what consumed or reconciled the missing ~21% weekly allowance and why the reset timestamp changed.

If the allowance was incorrectly consumed or reconciled, please:

1. restore the missing weekly usage / reset the affected allowance; and
2. because this also made a newly available 5-hour window unusable and interrupted paid development work, I would appreciate one month of ChatGPT Pro (20× Codex allowance) as compensation, or an equivalent amount of Codex usage credit if an account upgrade is not the appropriate remedy.

The primary goal is to get the accounting issue investigated and corrected; the compensation request reflects the lost paid usage and disruption if the bug is confirmed.

Additional information

• Plan: ChatGPT Plus
• Platform: Windows 11
• Local evidence: affected Codex rollout JSONL with exact quota snapshots
• Related tracker: #41220
• Similar recent reports: #44205, #44206

I prefer not to post the full rollout publicly because it contains private project activity, but I can provide a sanitized extract or the relevant telemetry records if maintainers need them.

No repository files, tests, or code entry points are identified. Begin with the server-side usage ledger and rate-limit bucket history for the account during 2026-09-09 14:52–19:08 CEST, using the rollout JSONL snapshots if provided. Done means identifying the missing allowance and reset change, or documenting why they occurred.