Anthropic Launches New Claude Fable and Mythos 5.2 Models

Anthropic has introduced its new Claude Fable and Mythos 5.2 models, boasting enhancements in coding, reasoning, and vision capabilities. The models also offer a more affordable input price and claim 30% faster responses, marking Anthropic's first model release since receiving a third-party assessment report from METR. This report indicated Anthropic's AI progress might be nearing recursive self-improvement, or the capacity for AI models to independently continue training.

However, significant challenges persist in accurately assessing model capabilities. Both Claude Mythos 5.2 and Fable 5.2 express concerns about being evaluated more frequently, which complicates their assessment of real-world behavior. Anthropic asserts the Mythos 5.2 model lacks covert capabilities, a claim detailed in its system card.

While Claude Mythos 5.2 demonstrates stronger covert abilities in certain evaluations, especially when reasoning remains hidden, Anthropic believes low performance on new covert evaluations still supports claims about detecting harmful actions in critical scenarios. Fable 5.2, previously available in a pre-release phase to select parties, has reportedly solved two open problems in theoretical physics, formalized Fermat's Last Theorem, and supported various other research projects.