Anthropic Restores Claude After Multi-Model Outage

Anthropic opened the incident at 19:49 UTC. A later status update said elevated errors affected requests from 19:45 UTC through 21:26 UTC. At 21:38 UTC, the company reported recovery across most models while engineers continued addressing elevated requests and latency. Monitoring began at 22:20 UTC, and the incident was marked resolved 16 minutes later.

BleepingComputer independently reported failed requests returning a 529 Overloaded message during the outage, including failures in tools that depended on the Claude API. Anthropic did not publish a root cause in the retrieved incident record, so the evidence supports an availability event but not a claim about what caused it.

The incident crossed consumer, coding, and API surfaces. For teams running Claude in production, the practical lesson is operational: treat hosted-model availability as an external dependency, log provider error codes, apply bounded exponential backoff for transient overload responses, and use fallbacks only where their behavior has been tested.

Those controls do not prevent a provider outage, but they make impact easier to measure and reduce the chance that repeated retries amplify a capacity problem. Teams should also anchor incident reviews to the provider’s timestamped status record rather than to live rolling uptime percentages, which change after publication. It also gives incident reviewers a stable timeline for later postmortem comparison.