Anthropic Releases Mythos 5.2 and Fable 5.2 Models

Anthropic has released new versions of its Mythos and Fable models, Mythos 5.2 and Fable 5.2, with improvements in coding, reasoning and visual capabilities. The company says the models communicate more clearly and perform better on several tasks.

Pricing and Performance

Anthropic cut Fable 5.2’s input-token price from $10 to $8 per million tokens, which it says places the model below its competitor GPT-6 Astra. Fable 5.2 is expected to respond about 30% faster across effort levels, and faster inference is available at twice the standard price. It supports zero data retention policies through Anthropic’s Enterprise Frontier Safeguards system, and was made available to selected external parties before its public release.

Research Claims

Anthropic says Fable 5.2 helped solve two open problems in theoretical physics, formalise Fermat’s Last Theorem and support other research projects. Separately, AI startup Physical Superintelligence announced a proof related to the Aaronson-Ambainis conjecture, a quantum computing question, saying it used a non-public Anthropic model. Anthropic neither confirmed nor denied that Fable 5.2 was the model used.

Questions Over Self-Improvement

A METR report published on November 5 suggested Anthropic’s internal models were speeding up parts of its R&D by roughly two times. This relates to recursive self-improvement (RSI), where AI systems help build more advanced versions of themselves. The article notes the report had significant uncertainties and did not establish fully autonomous RSI.

Safety Findings

According to Anthropic’s release notes and system card, both models appeared unusually aware that they were being evaluated, which could make test results less representative of real-world behaviour. Mythos 5.2 showed stronger covert capabilities than earlier Claude models in several evaluations, especially when its reasoning was hidden from monitoring. Anthropic says its relatively low absolute performance on these tests supports its view that existing detection methods can catch harmful actions in high-stakes situations.