ChatGPT Hit 75% Failure Rate During 3.5 Hour Outage

OpenAI's ChatGPT and API infrastructure experienced a catastrophic outage on November 21, 2026, with request failure rates peaking at 75% across both paid and free tiers. The incident lasted three and a half hours, from 1:48 to 5:21 PT, and impacted nearly every surface of the company's product stack, from conversational interfaces to backend API endpoints.
What Caused the Outage
According to OpenAI's post-incident writeup, the trigger was a scheduled maintenance window on a PostgreSQL database service backing the Primary API. This service handles user authentication and service-to-service authentication across the entire OpenAI stack. During the maintenance operation, both read replicas failed simultaneously, taking down the authentication layer that every other service depends on.
Because the Primary API is load-bearing infrastructure, the failure cascaded through dependent request paths. Chat Completions, Fine-tuning, Embeddings, Images, Batch, Audio, Moderations, Realtime, and Files endpoints all went down on the API side. ChatGPT users saw failures in Conversations, Login, Search, File uploads, and Voice mode. Attempts to revive the existing replicas were unsuccessful, forcing OpenAI to coordinate with its unnamed database provider to spin up two new replicas in the same region.
Availability Timeline and Impact
Officially launched on November 21, 2026, the outage began at 1:48 PT and persisted until 5:21 PT. For 211 minutes, the majority of requests to OpenAI's commercial and consumer products returned errors. The 75% peak failure rate represents one of the most severe public incidents in the company's operational history. Both free and paid ChatGPT users experienced identical impact, as the authentication failure sat upstream of all billing and tier logic.
Immediate Response and Long-Term Fixes
OpenAI has placed a moratorium on similar database maintenance operations until stronger safeguards are in place. The company retuned connection-churn settings on its PostgreSQL deployment and shipped a feature flag that allows operators to shift traffic away from replicas by percentage, giving teams more granular control during risky operations.
Longer-term remediation includes a pre-maintenance review process, holding adequate buffer capacity as hot standbys during maintenance windows, and building strategies to handle replica lag. These changes aim to prevent simultaneous replica failures and ensure at least one read path remains available during maintenance events.
What This Means
This incident underscores the operational complexity of running AI infrastructure at scale. A single point of failure in the authentication layer, compounded by simultaneous replica failures, can bring down an entire ecosystem of products. OpenAI's response, pausing maintenance and adding traffic-shifting controls, reflects the high stakes of uptime in a market where enterprise customers increasingly rely on API access for production workloads. The company's transparency in publishing the root cause and remediation steps sets a standard for incident communication in the AI industry.
on Emergent today





