For an hour on Tuesday, claude.ai, the desktop and mobile apps, Claude Code, Claude Cowork and parts of the Claude API were all unusable at once. By Anthropic’s own account the impact ran from 14:00 to 14:59 UTC. The status entry was closed at 16:27 UTC, more than ninety minutes after service came back.
How it unfolded
At 14:21 UTC the incident went up as “Investigating”: failed requests, errors loading or sending conversations, users being asked to sign in again. Seven minutes later came confirmation that the API was returning errors too. At 14:41 UTC Anthropic reported a first mitigation and sharply lower error rates.
Then it got messy. The 15:00 UTC update said the errors that started at 14:00 had been mitigated at 14:36 and that existing conversations mostly worked again. A second problem was blocking sign-in. Single sign-on and Sign in with Apple were unavailable, new chats, voice conversations and Claude Code and Cowork sessions would not start, purchases and file uploads failed. The advice was short and unambiguous: if you’re signed in, don’t sign out.
At 15:11 UTC most services were back as of 14:59 UTC. The same update carries the line that sticks: messages sent between 14:00 and 14:59 UTC may not have been saved.
The second failure was the worse one
The first wave hit requests. You notice that, you retry, eventually it goes through. The second wave hit authentication, and that is a different class of outage. Anyone who happened to have a live session could keep working. Anyone who had signed out, was starting fresh on a device, or sat behind an organisation’s SSO simply could not get in, however many times they tried.
For Claude Code and Cowork that was the hard part. Both start their sessions through that same sign-in, so they didn’t get slower, they didn’t launch at all. Purchases failing in the same window fits: this wasn’t the model, it was the account.
Anthropic hasn’t published a cause. A week earlier, on September 22, several models had thrown elevated errors, Opus 5, Mythos 5.1 and Fable 5.1 among them; that incident cleared after about an hour and a half and touched a much smaller slice of the product line.
Being already signed in was the difference
An hour of downtime is survivable. “Don’t sign out” is less so, because it shows how much hangs off a single path. Models, chats, code sessions, Cowork tasks and purchases all run through the same authentication, and when that goes, retrying doesn’t help and neither does switching models.
The practical takeaway is dull and worth writing down anyway. If you run Claude inside automation, hold an API key that doesn’t depend on interactive login, and don’t log out mid-session to troubleshoot. And if you typed something that mattered on Tuesday around midday, go check that it’s still there.