On November 21, OpenAI suffered its worst service outage of the year. Between 1:48 p.m. and 5:21 p.m. Pacific Time, request failure rates across ChatGPT and the OpenAI API peaked at 75%, and the incident lasted 211 minutes, hitting free and paid users alike. OpenAI published a post-incident write-up on its status page the same day, confirming the outage was triggered by routine database maintenance. All services were back to normal on November 22.
Both database replicas went down at once, and authentication fell first
The trigger was OpenAI's Primary API service, which handles user authentication and service-to-service authentication across the entire stack. Scheduled maintenance caused both PostgreSQL read replicas behind the service to fail at the same time; replica lag grew without bound, and every request path depending on authentication failed with it. On the API side, Chat Completions, Fine-tuning, Embeddings, Images, Batch, Audio, Moderations, Realtime and the Files endpoint were all affected; in ChatGPT, conversations, login, search, file uploads and voice mode failed one after another. At the peak, users filed thousands of error reports on Downdetector, and Codex showed instability too.
The real lesson is single-point dependence
What matters about this outage is not its length but its single point of failure. Once authentication goes down, chat, API, voice and file features that look independent all fall together. For developers and agent workflows that bet their whole business on OpenAI alone, three and a half hours of downtime is a full stop. OpenAI's write-up mentions no data loss or leak: this was an availability incident, not a security incident. The lesson is familiar — Microsoft 365 also went offline for 36 hours this month: Microsoft 365 back online after 36-hour outage: Outlook, Teams and Copilot recover. Outages at large cloud services are never a question of if, only of how often.
OpenAI promised three fixes
OpenAI has promised a moratorium on similar maintenance, retuned database connection settings to reduce connection churn, and shipped a feature flag that can shift traffic off read replicas by percentage. There is little users can do beyond subscribing to status-page notifications and keeping a backup provider for critical workloads. Whether the fixes are enough will only be proven at the next maintenance window.