Claude's fourth outage this month took down every surface at once
A ~3-hour wave of 529 errors hit Opus 5, Fable 5 and Opus 4.8 across API, Code and Cowork — the month's fourth. Time to wire a fallback before the next.
Copy markdown
Three hours of 529s, every model at once
Elevated “Overloaded” (529) errors hit Opus 5, Fable 5 and Opus 4.8 from 04:50 to 07:36 UTC — 08:50 to 11:36 in Dubai, right across the Gulf, India and Europe morning. Anthropic flagged the cause within ~30 minutes but took nearly three hours to fully clear it, and claude.ai, the API, Claude Code and Cowork all threw errors through the window.
The fourth wobble in a week
Claude also degraded on Aug 18, 19 and 20. The through-line is shared infrastructure: when one surface breaks, they break together, so you don’t get separate incidents — you get one unavailable dependency. A Claude-only stack has no graceful degradation when that happens.
Wire a fallback before the next one
Builders routing through OpenRouter or keeping a GPT-5.6/Codex fallback kept shipping through the outage. In Claude Code, ANTHROPIC_DEFAULT_MODEL plus custom base-URL/gateway support (v2.1.236 and up) makes swapping providers a one-line change — set it up now, not at 5 a.m. mid-incident.
Why a cheaper Plan B is trending
The top Hacker News thread this morning (460+ points) argues Anthropic’s flagship is bleeding daily users to cheaper, faster tools. Reliability plus last week’s GPT-5.6 Sol cut to $4/$20 is nudging more builders to a two-provider default — exactly the setup that rode out today’s window.