Azure failure likely took ChatGPT, Claude and Grok down together
A 3h37m East US region fault hit their APIs, not just the chat apps — Codex and Cursor stalled too. Gemini, on Google Cloud, sailed through untouched.
Copy markdown
What broke, and for how long
Starting around 15:07 UTC on Sept 3, ChatGPT, Claude and Grok all threw elevated errors within minutes of each other; Grok's own status page clocked 3h 37m of downtime before recovery near 19:40 UTC. Claude's Opus 5 was among the last models to come back.
The common fault line: Azure East US
Third-party monitor StatusGator traced it to ingress failures in Microsoft's Azure East US region, which all three vendors reportedly lean on. No company — nor Microsoft — has confirmed a root cause, so treat the Azure link as strongly reported, not official.
Your API calls failed, not just the chat tabs
This wasn't a consumer-app blip: the model API endpoints returned errors, so Codex and Cursor runs stalled mid-task alongside the web UIs. If your product calls these APIs directly, your users hit the outage the same moment you did.
Gemini stayed up — and that's the lesson
Gemini, running on Google Cloud instead of Azure, sailed through untouched. Provider diversity isn't redundancy if 'three independent labs' all sit on the same cloud region — the fault was in the infra beneath the models, not the models themselves.
Build this weekend: a fallback that actually fails over
Put a router (OpenRouter, LiteLLM, or your own) in front of your calls with a cross-cloud backup — Gemini or a self-hosted open-weight model — and rehearse it by assuming every primary AI API is down at once. That drill just stopped being hypothetical.