ChatGPT, Claude and Grok All Broke the Same Morning

Three of the largest hosted AI providers — OpenAI, Anthropic and xAI — all broke inside a three-hour window on Thursday, September 3.

All three are back online. What is worth keeping is the timeline, because it is the cleanest recent illustration of a dependency most teams have never actually tested.

The timeline

• 9:30AM ET — Grok began failing across Android, iOS and web. Prompting it on X returned: "This model is overloaded right now. Please try again shortly or pick a different model." xAI has since fixed the issue.
• ~11:00AM ET — ChatGPT started returning error messages, with OpenAI's status page reporting "elevated errors across ChatGPT and Codex."
• Around the same time — Anthropic's Claude chatbot, Claude Code and the Claude API went down. Anthropic technical staff member CJ Avilla attributed it to an "infrastructure issue" causing a partial outage across its services.
• ~12:15PM ET — Anthropic resolved its problem.

The OpenAI outage was not limited to chat. Per The Verge, it also affected logins, file uploads, voice mode, search, deep research and image generation — a reminder that "the model is down" and "the account layer is down" are different failure modes with different blast radii.

What is not known

There is no established common cause. The Verge is explicit: "It's not clear what went wrong at all three companies, or if the issues were somehow related," and the publication said it had reached out to OpenAI, xAI and Anthropic without an immediate response.

That matters for how you read the day. Simultaneity is not evidence of a shared dependency, and three separate incidents landing in one morning is a plausible coincidence. Anyone telling you a specific shared upstream failed is going beyond what has been reported. One piece of adjacent context: the incident happened while OpenAI was teasing the launch of Astra, its new AI model.

The part you control

This is analysis rather than reported fact, but the practical lesson is narrow and cheap to act on. If your product calls one provider's API on the critical path, Thursday was a live test of what your users saw — and most teams do not know the answer because they have never run it deliberately.

Three things worth checking this week:

• Do you fail over, or fail? A second provider behind the same interface only helps if the switch is automatic and someone has actually exercised it under load.
• Does auth degrade separately from inference? The OpenAI incident hit logins and uploads too. If your fallback path still needs the primary provider's account layer, it is not a fallback.
• What does a user see while you are down? Anthropic reported the problem resolved at around 12:15PM ET, after going down around the same time as ChatGPT. A queue with an honest message survives an hour of that; a spinner and a silent 500 do not.

Nothing here requires re-architecting around multi-provider abstractions today. It requires knowing which of your features are hard-wired to one vendor's uptime, and having that list written down before the next simultaneous morning.