OpenAI suffers 3.5-hour outage as twin auth-database replicas crash, peaking at 75% failure rate

OpenAI's November 21 outage lasted 211 minutes, and at its worst three out of four requests to ChatGPT and the developer API failed. That makes it one of the company's largest disruptions this year. It also lands in the same week OpenAI launched GPT-6.2 Astra and Sol, when traffic and developer attention were both unusually high.
The failure sat in a fairly mundane part of the stack. During scheduled maintenance, two PostgreSQL read replicas supporting the authentication gateway crashed simultaneously. Because auth sits in front of every model call, losing it effectively took the whole platform down, even though inference capacity itself was healthy. Recovery meant standing up replacement replicas and retuning the connection-pooling configuration that had amplified the load on the surviving infrastructure.
The competitive context matters. Enterprise buyers can now reach OpenAI's newest models through Amazon Bedrock and Microsoft 365 Copilot, as well as through OpenAI's own endpoints. An outage on the first-party API strengthens the argument for routing through a hyperscaler, or for keeping a Gemini or Claude fallback configured. Developer reaction centered on exactly that: agentic systems that call one provider for every step inherit that provider's single points of failure.
Watch for a full public postmortem, any SLA credits for API customers, and whether OpenAI re-architects auth for multi-region redundancy. Two questions remain open: why the redundant replicas failed together rather than independently, and whether maintenance windows will now be staggered.