September 22, 2026
Claude errored across API, Claude Code and app: 80 minutes to normal
For 80 minutes on the night of September 22, five Claude models and all four access points were returning errors at once.
On September 22 at 03:50 Moscow time, requests to Claude began returning errors.
It was not on your end. Anthropic itself detected an elevated error rate and kept the incident open for 80 minutes, until 05:10 Moscow time.
The outage hit four access points at once: claude.ai, Claude API, Claude Code, and Claude Cowork. Mythos 5.1, Fable 5.1, Opus 5, Fable 5, and Mythos 5 returned errors.
Anthropic began investigating seven minutes after the window opened, and 20 minutes later said it had found the cause. By 04:35 Moscow time, Mythos and Fable in both versions had returned to normal; Opus 5 took the longest to fix. At 05:11, successful-response rates had evened out across all affected models, and the incident was closed at 05:35.
Not a one-off. From September 1 to 22, the Claude status page recorded 12 incidents, including the same "Elevated errors for multiple models" on September 3. The entire visible feed since August 14 contains 28 entries.
What to do while they fix it. On Hacker News (72 points, 56 comments), practitioners settled on two workarounds: roll back to Opus 4.8, 4.6, or Sonnet, and restore a dropped Claude Code session with the command claude --resume.
In code, overload appears as 529 overloaded_error, while an internal error appears as 500 api_error. Official SDKs retry them automatically: twice by default, with increasing pauses and attention to the retry-after header, while the client's max-retries option changes the number of attempts.
status.claude.com itself sends incidents to subscribers by email, SMS, Slack, and Microsoft Teams, through a custom webhook, or via RSS. That arrives before failed requests tell you about the outage.
The Claude status page's September counter stopped at 12 incidents.
