Summary
On Sep 3, 2026, beginning at approximately 7:43 a.m. PDT, users experienced elevated errors while accessing ChatGPT and Codex. A routing configuration change prevented affected requests from reaching our services. The disruption was mitigated by approximately 8:20 a.m. PDT.
Impact
During this period:
Users experienced errors, including “not found” responses, across ChatGPT on web, mobile, and desktop, and Codex on the CLI and desktop app.
Some users also experienced errors when signing in, creating accounts, or accessing published sites.
Root Cause
Our routing system incorrectly interpreted a configuration update and stopped directing requests to available services. As routing was restored, returning traffic overloaded some services, prolonging recovery from the disruption.
Resolution
Engineers rolled back the configuration change to restore routing. The broad disruption was mitigated by approximately 8:20 a.m. PDT.
Prevention and Improvements
We have implemented or identified several improvements to reduce the likelihood and impact of similar incidents:
Keep valid routes available during configuration changes. Simplify routing dependencies so an inactive destination does not cause healthy destinations to be removed from service. Also improve fallback behavior if the system receives an invalid configuration.
Limit the reach of routing changes. Strengthen staged rollouts and evaluate health-check-based automatic rollbacks before changes reach more users.
Reduce errors during recovery. Improve how returning traffic is distributed so recovering services are not overloaded.
Detect and escalate routing failures sooner. We updated alerts to properly page outside business hours and are addressing monitoring delays and gaps in escalation.
We apologize for the disruption and are continuing to strengthen the reliability of the systems that support ChatGPT and Codex.
