Claude, ChatGPT and Grok broke the same morning

Anthropic said an infrastructure issue caused a partial outage across Claude.ai, Claude Code, Claude Cowork and the Claude API on September 3. ChatGPT and Grok failed for many users on the same day. All times below are converted to UTC.

MacRumors reported the overlap at 15:56 UTC. At that point OpenAI’s status page acknowledged issues across ChatGPT and Codex, Grok’s website said the service was having issues, and Anthropic’s status page showed elevated errors for Claude. All three companies said they were investigating. MacRumors later updated the story to say the chatbots were coming back online.

Claude’s trouble started first. Anthropic’s status history shows it began investigating at 13:26 UTC, about an hour and a quarter before the start of OpenAI’s failure, which the company put at around 14:43 UTC (7:43 a.m. Pacific). The two incidents overlapped, but they did not start together.

OpenAI said a routing error made ChatGPT and Codex unavailable for some users across platforms. It put a fix in place around 15:17 UTC and kept monitoring. OpenAI’s status page marked the incident resolved at 16:55 UTC and noted that some Codex remote control users may need to pair their mobile device again.

On Grok, xAI’s status pages listed a confirmed Models outage across Grok Web, Grok in X, iOS, Android and API regions, lasting roughly 3.5 hours. On X, Grok answered with “This model is overloaded right now,” The Verge reported.

Why all three platforms failed on the same day has not been publicly disclosed. Whether the timing reflects a shared cloud or network dependency or is a coincidence remains an open question.

Newsweek’s figures from user reports show how far the outage spread. Downdetector peaked above 35,000 user-submitted reports for OpenAI, roughly 80% of them tied to ChatGPT, and logged more than 1,000 reports each for Claude and Grok.

At 13:50 UTC, Anthropic’s page listed the affected models: Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8 and Opus 4.6. Recovery came in stages. A later update said the remaining models were back to their baseline error rate while Opus 4.8 and Opus 5 were still affected. Impact ended at 16:16 UTC.

Claude users have been here before. The August 5 incident ran 7 hours 8 minutes, the August 12 one lasted 4 hours 16 minutes, and the errors on August 18 were the third multi-model degradation in two weeks. Opus 5 was on the affected list in early August too, and on September 3 it was one of the two models that stayed affected longest.