Claude Errors on Sept. 22 Hit Three Models and Four Services
Anthropic posted an incident titled "Elevated errors for multiple models" on September 22, 2026. At 00:57 UTC, the company said it was investigating elevated errors on Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5. Anthropic Status
At 01:17 UTC, Anthropic said it had identified the cause and was working on a fix. Those first updates included no error rates, status codes, or regional detail.
The entry names four affected surfaces: claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork.
Anthropic's Claude Status history lists 55 incidents. A separate entry titled 'Elevated errors across many models' is marked as resolved.
Recent history gives context. MacRumors reported a service outage affecting multiple Claude models on July 6, 2026, resolved as of 09:13 UTC. MacRumors TechCrunch reported on March 2, 2026 that thousands of users reported problems accessing Claude services. TechCrunch On September 3, 2026, The Verge noted in an update that outages affecting ChatGPT, Claude, and Grok had been resolved. The Verge
Older reports extend the pattern. TechCrunch reported on September 10, 2025 an outage impacting APIs, Console, and Claude, and on June 4, 2024 that Claude and Perplexity began seeing issues on Tuesday morning. TechCrunch
The broader context here is operational, not mysterious. Brief updates with few details are normal in the first hour. The list covers direct use on claude.ai, programmatic inference through the API, where an app asks the model to generate a response, and assisted work in Code and Cowork. Like counters served by one kitchen, one backend problem can affect them all, so owners cannot assume one path will stay clean. In practice that means retries, growing queues, slow responses at the tail, and careful choices about failover to a backup. Logging the model identifier, pinning a version where allowed, and using the status page to pause less urgent work are basic hygiene. When Mythos 5.1, Fable 5.1, and Opus 5 are named together, switching model strings is unlikely to help. A simpler fallback, cached answers where acceptable, and clear messages to users matter more.
In my view, the encouraging part is speed of communication. Investigation at 00:57 UTC and cause identification at 01:17 UTC is a tight loop. Hosted AI will have bad hours. What sustains trust over the long arc is this cadence: name the affected models, name the affected surfaces, update quickly, then publish a fix. My kids grew up assuming the internet simply works. They are right to expect it, and the work that makes that expectation reasonable happens in incident threads like this one.
