ChatGPT, Claude and Grok degraded within half an hour of each other

Three of the largest chatbots started failing on Thursday inside the same half hour. Anthropic's Claude went first at about 14:30 UTC, OpenAI opened an incident at 14:58 UTC titled “Elevated errors across ChatGPT and Codex”, and xAI's Grok began returning an overload message on web, Android and iOS. None of the three has said what caused it, and none has said whether the failures are connected.

The timeline in UTC

• 14:30: Claude, Claude Code and the Anthropic API start throwing errors. The status page first lists Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8 and Opus 4.6, then narrows to Opus 4.8 and Opus 5 as the rest return to baseline error rates.
• 14:58: OpenAI opens its incident and marks it as investigating.
• 15:22: OpenAI applies a mitigation and moves to monitoring.
• 15:50: the mitigation holds, and OpenAI still reports partial system degradation. The incident remained open when this piece was written.

What stopped working

OpenAI's status page listed logins, file uploads, voice mode, search, deep research, image generation, Agent, Codex on the web and in the desktop app, the Codex API, the VS Code extension and connectors. Downdetector logged more than 36,000 reports against OpenAI around 14:54 UTC. Grok answered prompts on X with a line telling users the model was overloaded and suggesting they pick a different one.

“We're working on it but don't have an ETA yet.”

— CJ Avilla, technical staff, Anthropic, On X, via The Verge, 3 September 2026

Quote source: The Verge, 3 September 2026

Avilla described the Anthropic side as a partial outage caused by an infrastructure issue.

The unconfirmed part

Several outlets pointed at Cloudflare, which was having problems of its own and sits in front of a large share of the web. Neither OpenAI, xAI nor Anthropic has confirmed a shared cause, and The Verge said all three declined to comment in time for publication. The common-supplier theory is the obvious guess and remains unconfirmed.

The stakes changed this year

Chatbot downtime used to cost a person their afternoon. This year companies wired these models into production systems: 80% of the Fortune 500 report running AI agents, and Gartner lists inadequate risk controls among the reasons it expects over 40% of such projects to be cancelled by 2027. An agent that fails halfway through a task across four systems is a different incident from a chatbot that will not answer.

The timing is its own footnote. OpenAI spent this week promoting Astra, the model it says can build cyberattacks without human help, and spent Thursday unable to keep search and file uploads running.

For anyone with an API key, the question worth asking is whether the fallback path was ever tested. Thursday showed that a second provider does not help when it goes down in the same half hour.

This article is for informational purposes only and does not constitute investment advice.