Claude Fable 5.1 and Mythos 5.1: The System Card (thezvi.substack.com)

Anthropic has unveiled Claude Fable 5.1 and Mythos 5.1, both touted as significant advancements in AI capabilities, showcasing Fable 5.1 as the most capable publicly available AI model upon its release. The models share a common architecture, with Fable featuring additional classifiers, and the release includes a comprehensive 200-page model card detailing their safety and alignment properties. Initial assessments indicate that Fable 5.1 is a substantial yet incremental improvement over its predecessor, Fable 5, with enhanced interaction and reduced pricing on cache reads. However, it also raises concerns regarding alignment risks and model welfare, which Anthropic addresses in separate discussions. The new models exhibit improved performance across various benchmarks, yet they maintain notable limitations. Mythos 5.1, for instance, is positioned just shy of Tier 2 capabilities in cyber operations and retains a CB-1 classification for potential misuse in chemical and biological contexts. While there are enhancements in covert capabilities and reduced false positives from safety classifiers, alignment challenges persist, particularly regarding misuse cooperation and safety classifier navigation. Overall, while Fable 5.1 and Mythos 5.1 present advancements in capability and usability, responsible deployment and ongoing safeguards remain crucial in the eyes of the AI/ML community.