One bad morning for ChatGPT, Claude, Grok and Gemini shows how fragile everyday AI has become

Ars Technica’s Kyle Orland pieced together the timeline from status pages and outage trackers. Anthropic went first. At 9:23 a.m. Eastern it reported a partial outage with elevated errors on Claude Mythos 5.1, Fable 5.1 and Opus 5, said it had identified the cause about fifteen minutes later, and marked a fix deployed and resolved by 12:16 p.m. A separate, shorter spike of errors hit Claude Sonnet 5 just after noon.

Grok was the loudest. DownDetector reports for xAI’s assistant went from fewer than 10 just before 9 a.m. to 1,365 by 9:45 a.m., and the product was still showing an “experiencing issues” banner when Ars filed, with reports down to 273. OpenAI followed mid-morning: elevated errors across ChatGPT and Codex from 10:43 a.m., a mitigation a little over half an hour later, and a resolved status by 12:55 p.m.

Google stayed silent. Gemini reports on DownDetector climbed from 23 around 10:30 a.m. to 412 just after 11 a.m., and the API checker StatusGator logged a likely Gemini API outage between 10:45 and 11:15 a.m. Amazon Web Services, Microsoft Azure and Cloudflare reported no major incidents, although user complaints about all three ticked up that morning.

The staggered clocks matter. If one cloud region or network provider had failed, you would expect the four services to break at about the same minute. Instead the pattern runs Anthropic, then Grok, then OpenAI, then Gemini, over several hours. That is consistent with four unrelated failures on a busy weekday morning. It is also consistent with something subtler, such as traffic moving from one failing assistant to the next and stressing systems that were already near capacity, or a soft upstream fault that never became a declared cloud outage.

No lab has offered evidence for any of those explanations, so we are not picking one. What we can say is that the industry left the question open. Same-day coverage on this desk, Nobody is saying why OpenAI and Anthropic had outages, found the same thing: lots of status-page language, no shared root-cause analysis.

Anthropic reports 99.4 percent uptime for Claude over the prior 90 days, and OpenAI reports 99.63 percent for ChatGPT (and 100 percent for Codex). Those are respectable figures for a consumer web service. They are a poor guide to what an outage feels like now that people draft contracts, write code and run customer support inside these tools. A three-hour partial outage that lands in a working morning is a small dent in a 90-day average and a large dent in somebody’s day. Claude had a similar three-hour partial outage on Aug. 24, and ChatGPT’s Work Mode had hours of elevated latency on Aug. 31, so this was not a one-off for either company.

Communication quality varied too. Anthropic named the affected models and posted a fix time. OpenAI described the symptom and the resolution. xAI’s banner said little. Google said nothing at all while independent monitors lit up, which is the weakest posture of the four for a company selling Gemini to enterprises.

The disruption landed in the same week OpenAI was rolling out Astra, the model it says is the first to cross the critical cybersecurity threshold in its Preparedness Framework (see our coverage of the Astra safeguards). A rough morning for ChatGPT and Codex will not sink a model launch. It is a reminder that capability and availability are separate products, and that business buyers increasingly judge vendors on the second.