Elevated errors for multiple models – Resolved
summary
Anthropic experienced a service disruption affecting multiple Claude AI models, including Mythos and Fable, leading to elevated error rates for users. The issue impacted several platforms including the web interface, API, and specialized tools like Claude Code. Engineers identified the cause and implemented a fix to restore normal success rates across all affected models. The total duration of the impact lasted approximately ninety minutes.
key points
- The outage primarily affected Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5, causing high failure rates for requests.
- Service disruptions were felt across multiple access points, including the claude.ai website, the Anthropic API, Claude Code, and Claude Cowork.
- The technical team successfully identified the root cause and restored full functionality between 5:50pm and 7:10pm PT on September 22, 2026.
who should read it
Users of Claude AI services and developers relying on the Anthropic API who need to confirm the cause of recent service errors.
discussion positive
consensus
Users agree that the most effective way to manage LLM workflows is through a 'human-in-the-loop' approach, specifically using multiple models and harnesses for different tasks like specification, implementation, and review. There is a general consensus that state should be managed locally—via the file system or Git—to make switching between models seamless. While some argue that specialized switching tools are unnecessary, others find custom routers or handoff files invaluable for maintaining continuity.
top insight
The practice of 'adversarial review'—using one high-end model to critique the architectural design produced by another—revealed significant flaws that a single model missed. However, the user warns that this process can lead to a 'minced up concept' if not carefully moderated by a human, as models can become overly conservative or nitpicky to avoid 'AI psychosis'.
Generated by an LLM from the article and thread. Check the sources before relying on it.