ChatGPT and Claude downtime: 8 and 13 hours in the past 90 days

Three of the biggest AI services went down within a few hours of each other on Wednesday, September 3. OpenAI logged an incident titled "Elevated errors across ChatGPT and Codex" at 14:58 UTC, affecting 15 ChatGPT components and four Codex components. The company marked it resolved at 16:55 UTC.

Anthropic recorded two separate incidents the same day. The first hit Claude Sonnet 5 from 12:37 to 12:56 UTC. The second ran from 13:26 to 16:16 UTC and covered, by Anthropic's own list, Mythos and Fable 5.1, Mythos and Fable 5, Opus 5, Opus 4.8 and Opus 4.6. xAI logged an incident for Grok as well. Google's Gemini was the only one of the four major services without an official outage notice, even though Downdetector still collected around 500 reports for it. ChatGPT peaked at more than 35,000 reports in the US, Claude at roughly 1,400 and Grok at roughly 1,200.

All of that is a snapshot, and it is already over. The more interesting part starts when you step away from the single day and look at what the providers publish about their own reliability. Those numbers are public, and almost nobody reads them.

The numbers sit on the providers' own status pages

OpenAI reports 99.64 percent availability for ChatGPT on status.openai.com, measured over a rolling 90-day window. The APIs do considerably better at 99.94 percent, and Codex sits at 100 percent. Anthropic publishes 99.4 percent for claude.ai on status.claude.com over the same period, 99.5 percent for the API, 99.44 percent for Claude Code and 99.43 percent for Claude Cowork. Both companies note that these are aggregate figures across all tiers and models, so individual experience can differ.

Percentages with two decimals do not tell you much, so here are the same values as time. Over 90 days, 99.64 percent means roughly 467 minutes without service, just under eight hours. The 99.4 percent for claude.ai works out to about 778 minutes, a little over 13 hours. That is a full working day plus overtime, spread across a quarter.

The incident history says the same thing. OpenAI's public list carries 36 individually logged incidents for July, 19 for August and another five for the first three days of September. Wednesday was not an outlier, it was the day several providers failed at once and people noticed.

Nobody involved has named a cause

A lot of coverage points to a failure in the Azure East US region, where ChatGPT, Claude and Grok all draw compute, but that link is not documented. Microsoft's official Azure status history, where it collects its Post Incident Reviews, carries no entry for East US on that day. The most recent report published there dates from July 23 and concerns the West US region.

Neither OpenAI nor Anthropic gives a reason in its incident notes. That stands out at Anthropic in particular, because the company had been explicit only days earlier. On August 28 it stated that it had identified "an issue with an upstream cloud provider" affecting Claude Cowork and Claude Code on the web. For Wednesday, no such attribution appears. Anyone presenting the Azure explanation as settled is going beyond what is on the record.

OpenAI did add one practical note. If you use Codex through the mobile remote control, you may need to pair your device again after this incident.

How to check whether the problem is yours

When a chatbot suddenly throws errors, the first instinct is to look for the fault at your end, switch browsers, sign out and back in. Three addresses save you that. OpenAI keeps its current state and a full incident history at status.openai.com, Anthropic at status.claude.com, xAI at status.x.ai. Gemini appears in the Google Workspace Status Dashboard.

These pages come with a catch. The providers maintain them themselves, and they regularly lag behind what users are seeing. On Wednesday the reports on Downdetector spiked before the official entries appeared. Checking both is a reliable way to separate your problem from theirs.

A second subscription only helps so much

The obvious move is to sign up with two providers. Wednesday shows why that is not a safe plan. ChatGPT, Claude and Grok were all degraded in the same window while Gemini kept running. Which service sits on which infrastructure, and where two providers share a dependency, is close to impossible to tell from the outside. Two subscriptions protect you only if they happen to stand on different legs.

A model that runs on your own hardware without an internet connection is more dependable. Those models are now small enough for ordinary devices. We have shown how a language model fits on an iPhone, and Meta followed up with Muse Glimmer, a free option from 24 GB of graphics memory. They do not match the quality of the large cloud models. For the case where your provider is simply unreachable, they are still worth more than a second monthly bill.

Outages at Anthropic are nothing new either, as we reported in late August, when several Claude models threw errors at the same time.