Cross-Outage: ChatGPT, Claude, Grok and Gemini All Broke at Once

An American morning full of holes

Anthropic is the first to put its hand up. At 9:23 a.m. Eastern on Thursday, September 3, the company declares a partial outage: elevated error rates on requests to Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5. Fifteen minutes later it says it has found the cause; a fix ships and the incident is closed at 12:16 p.m. A second report, just after noon, flags the same symptom on Claude Sonnet 5.

OpenAI walks on stage at 10:43, with elevated errors on ChatGPT and Codex plus degraded performance, resolved at 12:55 p.m. — a good half hour after a mitigation went in. Grok, meanwhile, was still showing a notice saying the model was having problems when Ars Technica published its piece: DownDetector reports climbed from under ten just before 9 a.m. to 1,365 at 9:45, then fell back to 273. Gemini traced the same curve more discreetly — 23 reports around 10:30, 412 just after 11 — and API monitoring outfit StatusGator logged a probable API outage between 10:45 and 11:15.

The shared supplier nobody could find

Four competitors tripping over the same doorstep in the same window, and the instinct kicks in straight away: look one floor down, at the company that never appears in anyone’s status post because its logo isn’t on the homepage. Except the usual suspects filed nothing. Amazon Web Services, Microsoft Azure and Cloudflare were all reporting no major incidents at press time — even though their own DownDetector counters ticked up that morning too, which, when four assistants are throwing errors at once, may simply mean users blaming anything that owns a status page.

The shared-dependency theory is a lovely story. Nothing made public actually backs it. None of the four has disclosed a root cause. Anthropic says it identified its own; what it was, we still don’t know.

Rare, but not impossible

The uptime numbers are a reminder that none of these services runs clean: 99.4% for Claude over the last 90 days, 99.63% for ChatGPT and 100% for Codex across the same period. Claude had already taken a three-hour partial outage on August 24, and ChatGPT’s Work Mode ran several hours of high latency on August 31. Everyone gets a bad day; what’s practically unheard of, per Ars Technica, is that they synchronised. Correlation still doesn’t manufacture a common cause: 99.4% uptime leaves you about a dozen hours of downtime a quarter to line up.

What you really have in common is your visibility

Google has acknowledged nothing at all. So everything we have on Gemini comes from outside telemetry: a tally of complaints from the general public, and a third party poking the API from the pavement. That may be the most instructive part of Thursday morning. You can spread your bets across providers, write a clean failover between four models, and still find yourself diagnosing your own production by hammering refresh on a consumer complaints page.

The only firmly documented shared dependency that Thursday wasn’t a datacenter — it was a complaints counter. Have a backup plan, and make sure it isn’t another chatbot.