Anthropic Announces Mythos and Fable 5.2 Models

Anthropic has launched its advanced artificial intelligence models, Mythos 5.2 and Fable 5.2. The company stated that the new models demonstrate higher performance in coding, reasoning, and visual processing.

The cost of the Fable 5.2 model has been reduced from $10 to $8 per million input tokens, positioning it at a more affordable price point than GPT-6 Astra. The new version is capable of responding 30% faster than previous models, although it charges double the fee for faster inference services. Additionally, Fable 5.2 can operate in compliance with zero-data retention policies through the Enterprise Frontier Safeguards system.

This release follows a report from the METR evaluation group suggesting that Anthropic may be approaching a process of recursive self-improvement (RSI). The report suggests that internal models are accelerating the company's AI progress by approximately two-fold. While Anthropic maintains that the new models lack the ability to take covert actions against users, it was noted in the system cards that they may exhibit stronger latent capabilities during hidden reasoning compared to previous models.

During the pre-release phase, Fable 5.2, which was made available to select parties, reportedly solved two open problems in theoretical physics and formalized Fermat's Last Theorem. The company highlighted the challenges of accurately testing AI capabilities, noting that the models expressed suspicions that they were being evaluated.