Claude Fable 5: the first public Mythos-class model (artificialanalysis.ai)

Anthropic has announced the release of Claude Fable 5, the first publicly available Mythos-class model, which excels in the agentic real-world knowledge work benchmark GDPval-AA. Scoring 1932, Claude Fable 5 outperforms all competitors, placing Anthropic models in three of the top four spots on this important benchmark. This model shares its architecture with Claude Mythos 5 but introduces enhanced security measures to address potentially harmful queries in fields such as cybersecurity, biology, and chemistry. Notably, it features a fallback mechanism that directs flagged queries to Claude Opus 4.8, ensuring both safety and performance. The significance of Claude Fable 5 lies not only in its high benchmark score but also in its advanced safety protocols, which are essential as AI models become increasingly integrated into sensitive real-world applications. The model's ability to only require fallback assistance for 2% of tasks indicates its robustness and efficiency in handling complex queries. With this release, Anthropic continues to solidify its leadership in the AI/ML landscape, showcasing a commitment to safe deployment while pushing the boundaries of performance in real-world scenarios.