Skip to header Skip to main navigation Skip to main content Skip to footer
Scott Lawson
Your next upgrade lives here

OpenAI Outage Failed Up to 75 Percent of Requests After Database Maintenance Crashed Both Auth Replicas

The office building at 1515 Third Street in San Francisco, headquarters of OpenAI, seen from across the street on a clear day.

A routine database maintenance job turned into one of OpenAI's worst outages of the year on Saturday, November 21, knocking out much of ChatGPT and the company's developer API for three and a half hours.

Between 1:48 and 5:21 p.m. Pacific time, request failure rates peaked at 75 percent across both paid and free ChatGPT and the API, according to a post-incident write-up OpenAI published on its status page the same day. The incident lasted 211 minutes in total, and all systems were reported operational again by Sunday morning, November 22.

The trigger, in OpenAI's own words, was maintenance on a Postgres database service used by its Primary API service. That Primary API service handles authentication across the entire OpenAI stack, which means it checks who a user or application is before anything else can happen. During the maintenance, both of the database's read replicas failed at the same time, and authentication went with them.

That single failure explains why the outage was so wide. On the API side, it swept across Chat Completions, Fine-tuning, Embeddings, Images, Batch, Audio, Moderations, Realtime, and the Files endpoint. Inside ChatGPT, conversations, login, search, file uploads, and voice mode were all affected. When the system that verifies identity stops working, even healthy services behind it become unreachable.

There was no sign of a cyberattack, a traffic surge, or a failed product launch in OpenAI's account. The cause was scheduled work on an ordinary, if critical, piece of infrastructure. That ordinariness is part of the lesson. Modern AI products are often described in terms of models, chips, and data centers, but they also rest on conventional databases and authentication services, and a failure in one of those familiar layers can stop everything above it.

The timing matters for customers as well. Many businesses now treat ChatGPT and the OpenAI API as infrastructure, building customer support, coding, search, and agent workflows on top of them. A 211-minute outage on a Saturday afternoon is a reminder of what a single-provider stack means in practice: when the provider's authentication fails, the customer's own systems may have nothing to fall back on, no matter how well they are built.

OpenAI deserves credit for publishing a candid, detailed write-up the same day, naming the maintenance, the database, and the simultaneous replica failure rather than hiding behind vague language about technical difficulties. That kind of transparency helps developers plan, and it gives the wider industry a concrete case to learn from about redundancy. Two replicas only protect a service if they do not share the same failure during the same maintenance window.

For now, service is restored and the status page shows no ongoing incident. The larger question the outage leaves behind is not whether OpenAI can fix a database problem. It is how many of the companies building on AI have a plan for the next afternoon when the login system, not the model, is the thing that goes down.

Photo: The office building at 1515 Third Street in San Francisco, headquarters of OpenAI, photographed by Coolcaesar, licensed under CC BY 4.0 via Wikimedia Commons.

Artificial Intelligence
Technology
OpenAI
Big Tech
Generative AI
Cloud Computing

Copyright © 2026 Rocky Mountain Madman LLC - All rights reserved

Developed and Designed by Rocky Mountain Madman LLC