OpenAI has explained how a routine database maintenance job turned into one of its worst outages of the year, knocking out three-quarters of ChatGPT and API requests for three and a half hours on November 21.

Between 1:48 and 5:21 p.m. Pacific time, request failure rates peaked at 75% across both paid and free ChatGPT and the company's API, according to a post-incident write-up published on OpenAI's status page the same day. The incident lasted 211 minutes in total.

The outage hit almost everything OpenAI runs. On the API side, Chat Completions, Fine-tuning, Embeddings, Images, Batch, Audio, Moderations, Realtime and the Files endpoint were all affected. Inside ChatGPT, conversations, login, search, file uploads and voice mode all degraded or failed outright.

Two replicas, one maintenance window

The trigger, in OpenAI's own words, was maintenance on a PostgresDB service used by its Primary API service, the system that handles authentication across the entire OpenAI stack. The maintenance caused both of the service's read replicas to fail at the same time.

With both replicas down, the authentication layer could not serve requests, and the failure cascaded across ChatGPT and the API. OpenAI said the incident was a reminder of how much of its stack funnels through a single authentication path, and it published the write-up the same day, a level of transparency that is still unusual among large infrastructure providers.

Users reported the problems in large numbers on Downdetector through the morning, with ChatGPT accounting for roughly 80% of complaints, alongside smaller clusters pointing at Codex, the app and login functions.

A fragile single-provider stack

The outage lands at a moment when businesses are building ever more of their daily work on top of a single AI provider. When ChatGPT and the API go down together, there is no fallback inside OpenAI's own stack, and companies that route customer support, coding and document work through it simply stop.

That fragility is exactly what the incident exposed. A single maintenance operation on one database service took down authentication for the whole platform, and the redundancy that was supposed to protect it, the second read replica, failed in the same window.

OpenAI has not said whether it will change how it runs maintenance on the Primary API service, but the shape of the failure points to the obvious fix: maintenance that cannot take down both replicas at once, and an authentication path that does not depend on a single database service.

Service restored, questions remain

All systems were reported operational again by the morning of November 22, and OpenAI's status page showed no ongoing incident. The company's most recent earlier incident, elevated error rates for GPT-6 Luna API requests on November 19, lasted under an hour.

For users, the practical lesson is familiar: even the largest AI providers go down, and when they do, they can go down hard. Businesses that depend on ChatGPT or the API for time-sensitive work will be weighing backup providers and offline fallbacks after a morning when three out of every four requests failed.

By Hamza Faisal