Anthropic announced the release of Claude Fable 5, the first publicly available Mythos-class AI model. According to Artificial Analysis, the model secured the #1 position on the GDPval-AA benchmark, a standard designed to measure agentic real-world knowledge work, with a score of 1932. Claude Fable 5 utilizes the same underlying architecture as Claude Mythos 5 but incorporates additional security guardrails to monitor for potentially harmful queries in fields such as cybersecurity, biology, and chemistry.

The release features a new fallback mechanism that automatically redirects flagged queries to Claude Opus 4.8. During testing, Claude Fable 5 utilized this fallback for 2% of GDPval-AA tasks. Anthropic estimates that this routing process occurs in fewer than 5% of user sessions on average, ensuring consistent model performance.

Anthropic announced the release of Claude Fable 5, the first publicly available Mythos-class AI model. According to Artificial Analysis, the model secured the #1 position on the GDPval-AA benchmark, a standard designed to measure agentic real-world knowledge work, with a score of 1932. Claude Fable 5 utilizes the same underlying architecture as Claude Mythos 5 but incorporates additional security guardrails to monitor for potentially harmful queries in fields such as cybersecurity, biology, and chemistry.

The release features a new fallback mechanism that automatically redirects flagged queries to Claude Opus 4.8. During testing, Claude Fable 5 utilized this fallback for 2% of GDPval-AA tasks. Anthropic estimates that this routing process occurs in fewer than 5% of user sessions on average, ensuring consistent model performance.