OpenAI investigates Codex banked-usage reset problems

• Codex users saw remaining weekly capacity collapse without active work: bridgemindai reported a 70% to zero drop, while nicdunz reported 56% to zero with no threads running.
• The failure involved banked resets that did not fully apply in ChatGPT Work and Codex, according to thsottiaux's explanation.
• OpenAI identified the issue, applied mitigation, and said service should return to normal, as reach_vb's recovery post put it.
• A later user report described a changed scheduled reset date after recovery, as koltregaskes posted.

OpenAI's incident log records a 25-minute path from investigation to resolution. Its banked-reset explainer defines the feature as a one-time stored reset, not purchased credit or a permanent plan increase. A contemporaneous Developer Community report described a separate user returning after three hours to find roughly 75% remaining capacity reduced to zero.

The reports were not uniform resets. They described counters falling to zero, falling partway, or returning at different levels.

• daniel_mac8 said a meter fell from 85% to 40% while their laptop was asleep.
• nicdunz reported a 56% to zero change with no active threads.
• Hangsiin said their account came back at 21%, while a colleague's returned to 100%.

A rate-limit counter is scheduling infrastructure for long agentic runs. The failure temporarily made the displayed balance a poor signal of available work.

OpenAI's help article says a banked reset is saved to an account until use or expiry, and does not raise the plan's standing limit.

thsottiaux said the incident occurred when a banked reset was used in ChatGPT Work or Codex and did not fully apply. The stated remedy was another reset for every account that used one during the affected window, plus an apology email.

The official status history records four updates on September 9:

1. 1:29 PM: OpenAI began investigating unexpected usage-limit resets affecting some Codex users.
2. 1:40 PM: The issue was identified and a fix was in progress.
3. 1:45 PM: Mitigation was applied and recovery was being monitored.
4. 1:54 PM: The page marked all impacted services fully recovered.

A direct reply from dkundel called the change a mistake on OpenAI's end and said recovery should be underway.

Recovery did not end every visible oddity in the reports. koltregaskes posted that their weekly reset day had moved two days later, and the same user later said they had used a banked reset before the date changed.

The public incident log ends with the recovery notice and does not explain whether the reported reset-date change was part of the mitigation or a separate account-state effect.