Claude Opus 5 is throwing 529s again — fall back to 4.8 or Sonnet
A second Opus 5 incident in a day hits Claude.ai, the API, Cowork and Claude Code with overload errors — worst under peak load. Console stays up.
Copy markdown
What's failing, and where
Opus 5 is returning timeouts, 529 “Overloaded” and internal-server errors across Claude.ai, the API, Claude Code and Cowork; Console and Claude for Government stay up. Anthropic opened one incident at 08:16 UTC and a second at 11:27 UTC.
It's a load wobble, not a bug you can fix
Builders report 529s persisting after the status page flips green, plus quality degradation at peak hours — self-contradictory answers and mangled explanations even when the underlying work is right. Off-peak requests fare noticeably better.
Your fallback is one flag away
Opus 4.8 and Sonnet 5 are unaffected. In Claude Code, switch with /model or /fast (it now maps to Opus 5 and Opus 4.8); on the API, pin the claude-opus-4-8 or Sonnet 5 model ID until Opus 5 stabilizes.
Wire in retry-and-failover before load climbs
529 is a retryable “overloaded” signal, not a hard failure — add exponential backoff plus a second-model failover route so a capacity spike degrades gracefully instead of dropping requests. It matters more with Opus 4.1 retiring Aug 5 and funneling traffic onto Opus 5.