Claude Was Down for 36 Minutes, Not 42. It Started as an Authentication Failure.

Anthropic logged a service disruption on 16 August that began as a sign-in failure and spread into degraded performance across five products. The status page opens the incident at 21:58 UTC and closes it at 22:34 UTC: 36 minutes.

The timeline, as published

Four updates, all stamped in UTC. At 21:58: "We are investigating an issue with users authenticating to claude.ai, Claude Code, and Claude Cowork." At 22:02 the scope widened to "degraded performance affecting claude.ai, platform.claude.com, the Claude API, Claude Code, and Claude Cowork". At 22:22 a fix was deployed and monitored. At 22:34 the incident was marked resolved.

What the common framing gets wrong

Two numbers and one scope claim are circulating incorrectly. The duration now repeated in aggregation is 42 minutes. It derives from a write-up filed at 22:40 UTC — while the incident was still open — and treats the moment of that reporter's observation as the moment of recovery. Anthropic's own page closes it at 22:34. Use the first-party timestamp.

Early coverage also reported that the Claude API and Console were unaffected. The final resolution message names platform.claude.com and the Claude API among the affected components. The API was degraded, not adjacent to the problem.

And "outage" overstates the first phase. For the opening four minutes the fault was described specifically as authentication — users unable to sign in, rather than a service that had stopped serving. The distinction matters because the two failure modes have different blast radii: an authenticated long-running session may continue while new sign-ins fail.

What has not been said

No root cause has been published. There is no post-mortem, no contributing-factor note and no statement on whether the authentication failure caused the wider degradation or was the first symptom of it. Any account asserting a cause is speculating.

Why 36 minutes is not a small number

The affected list is the point. Claude Code and Claude Cowork sit inside paid engineering workflows, not chat windows. A failure in the authentication layer does not degrade gracefully there — it stops work at customer sites, on customer deadlines, with no local fallback. Thirty-six minutes of that is a different event from thirty-six minutes of a consumer assistant being slow, and it is why the absence of a published cause is the more durable story.