Codex fixes usage overcounting with one extra banked reset and auto-review rollback

• OpenAI gave Codex users another full reset after saying several smaller issues, not one single bug, were compounding usage for some accounts, according to OpenAI's fix breakdown.
• The concrete fixes were an auto-review rollback, less subagent work, and repairs for background suggestions that could run twice or retry too often, per the fix breakdown.
• OpenAI also changed reporting so auto-review no longer shows up as GPT‑5.4 usage, and failed or rate-limited requests no longer appear in turn graphs, according to the same post.
• Users get one extra reset credit in their bank within 24 hours, while a prior hard reset post notes some accounts had already stacked up to three banked resets.

You can check your reset bank with the local auth flow prompt, and OpenAI's earlier Sunday update makes clear the team had already gone into a war room before the second reset. One useful detail buried in replies is that background usage can also come from automations, memory, and suggested prompts, with memory and suggestions normally around 1% each.

The interesting part of OpenAI's breakdown is that the drain came from stacked behavior in the background layer, not from a single runaway meter.

It names three separate causes:

• Auto-review had become more proactive.
• Another change was triggering more subagent work.
• Background suggestions could run twice, or retry too frequently after failures.

That lines up with the earlier Sunday update, where OpenAI said the Codex team was combing through logs before it had a root cause.

OpenAI reset everyone's limits once during the investigation, then said in the follow-up fix post that it would fully reset limits again and add one extra banked reset within the next 24 hours.

The reporting changes are as notable as the usage fixes:

• Auto-review had been showing up as GPT‑5.4 usage.
• Failed or rate-limited requests were still being shown as turns.
• Historical charts may keep old auto-review data under GPT‑5.4 because past turn data was not relabeled.

In a follow-up reply, OpenAI added that people who had already burned a prior reset shortly before the hard reset would get more manual resets after the investigation concluded.

The extra credit landed awkwardly enough that users immediately started comparing notes on what a "reset bank" even showed. aibuilderclub_ posted a prompt that reads ~/.codex/auth.json locally and queries the reset-credit endpoint, with an example response showing available_count: 4.

Not everyone could see the bank in product UI. In one thread, a user reply said they had never seen a reset bank at all, before a later reply suggested they were on Codex for OSS rather than the main Codex app.

One reply from OpenAI adds a detail that never made it into the longer incident posts. According to Thomas Sotiaux's background-usage reply, three features can consume usage in the background:

• automations
• memory
• suggested prompts

Sotiaux said memory and suggested prompts usually cost around 1% of the weekly limit each under normal conditions, and all three can be disabled in settings. That gives the weekend drain reports a more concrete baseline than the main reset posts did.