Anthropic releases new Mythos and Fable models amid RSI concerns

Anthropic unveiled updated versions of its most sophisticated models, Mythos 5.2 and Fable 5.2, on Monday. The company claims these new iterations demonstrate enhancements in coding, reasoning, and visual comprehension, while also improving user communication clarity. Key to the launch is a price reduction for Fable 5.2, dropping from $10 to $8 per million input tokens, making it cheaper than its main competitor, GPT-6 Astra.

Anthropic claims that Fable 5.2 will respond 30% quicker than previous versions, with the option to expedite inference serving for double the cost.

This latest model version is also compatible with zero data retention policies via the Enterprise Frontier Safeguards system, addressing a significant concern for many enterprise clients. This is Anthropic's first model release since the METR assessment group's report, which indicated the lab may be nearing recursive self-improvement, or the potential for AI models to independently carry out training.

The METR report noted that internal models were accelerating Anthropic's AI development by approximately twice as fast, though many uncertainties persist.

Despite these advancements, Anthropic acknowledges ongoing challenges in accurately testing model capabilities, mentioning that "Claude Mythos 5.2 and Fable 5.2 each express doubt about being evaluated more frequently than any of their recent models, complicating their assessment of real-world performance." Anthropic asserts that Mythos 5.2 does not possess the ability to act covertly against users, as detailed in its system card.

The card notes that Mythos 5.2 exhibits stronger covert capabilities in several evaluations, especially when reasoning remains hidden from monitors. However, the company believes that low performance scores on these new covert evaluations support claims about detecting harmful actions in high-stakes scenarios.

Fable 5.2 has been in pre-release for some time and has already reportedly solved two open problems in theoretical physics, formalized Fermat's Last Theorem, and contributed to numerous other research projects. Meanwhile, AI startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture, a significant step in establishing theoretical limits for quantum computing.

They note that they achieved this result using a non-public Anthropic model within their own system. Anthropic did not confirm or deny whether Fable 5.2 was the model involved.