OpenAI resets Codex usage limits after outage

OpenAI's coding agent Codex stopped working for 56 minutes on September 25 (UTC). The OpenAI status page first reported the outage at 22:58 and said at 23:54 that all affected services had recovered. Right after recovery, Tibo, who leads Codex, said on X that usage limits would be reset for all paid users of Codex and ChatGPT Work.

The severity was at the top of the scale. The status page classified the incident as a full outage rather than a partial one. According to reports, the affected components included Codex web, the Codex API, the command-line tool (CLI) and the VS Code extension, so there was no way to use Codex at all. It happened at 3:58 p.m. on a Friday in California and around midnight in London.

The recovery left six updates on the status page. At 22:58 OpenAI wrote that users were experiencing a Codex outage and that it had identified the internal issue and was heading toward mitigation. At 23:03 it confirmed elevated errors, and at 23:19 it offered a workaround: "Login via API key will unblock access at this time." At 23:34 it said it had identified the root cause, at 23:45 it applied the mitigation and monitored recovery, and at 23:54 it declared the incident resolved. Recovery took 20 minutes after the cause was found.

Users saw authentication errors. A post on the OpenAI developer forum shows that both the Codex interface and the CLI of a user signed in with a ChatGPT account returned "401 Unauthorized: Incorrect API key provided." The author wrote that they had not entered any API key of their own, and that updating the ChatGPT macOS app, restarting the Mac, regenerating the auth file and switching networks all produced the same error. The CLI version was 0.158.0-alpha.2, and after the WebSocket connection failed, the fallback to HTTPS returned the same 401. According to reports, Codex threads on the web broke off with a message stream error while regular ChatGPT chat kept answering, and issue 48234 reporting a total outage was opened on the Codex GitHub repository before being closed as a duplicate of 48237.

The cause has not yet been disclosed. The status page only said the root cause was identified at 23:34 and did not describe it. Reports suggested, based on the burst of authentication errors and the API key workaround, that the problem sat in the authentication step between clients and the service, but OpenAI has not confirmed this.

Compensation came 13 minutes after recovery. Tibo had posted a brief lament during the outage, and at 00:07 UTC on September 26 he quoted it and wrote, "we're back in action and we'll reset usage limits for all paid users across codex and ChatGPT work." He then apologized, "sorry about the brief disruption!", and added in parentheses, "and yes we have a special spare codex when things are down to help us out." The post passed 2.5 million views.

Resetting usage limits has become a familiar OpenAI apology. It is not a cash refund or a subscription extension but a usage credit that refills weekly quota already spent. Tibo's post named only paid users. According to reports, OpenAI also reset usage on all paid plans on August 25 after repeated reports that quotas were draining faster than expected. METAL has reported that Anthropic introduced a wrap-up feature for Claude Code's five-hour limit, and this incident again showed that the resource coding agent companies handle most carefully is the usage limit.

The status page METAL checked also carries a caveat. OpenAI says its availability metrics are aggregated across all tiers, models and error types, so individual customer availability may vary by subscription tier, model and API features in use. The API key login offered as a workaround is not free either. According to reports, API usage is billed per token, outside the subscription quota.

From a reporter's perspective, the question these 56 minutes leave is not compensation but explanation. OpenAI has been pushing this agent into other companies' products, including by opening the Codex harness through its API, and teams that hand it long jobs such as migrations or large refactors lose both quota and work state in an hour of downtime. OpenAI had a spare Codex for when the service stops, but customers only had whatever backup they had prepared themselves. Turning the 23:34 note that the root cause was found into a public explanation would restore more trust than a limit reset.