Anthropic Releases Fable and Mythos 5.2 Models

Anthropic says the new models produce clearer responses and can generate outputs faster than their 5.1 predecessors. Fable 5.2 is aimed at demanding software engineering work, including coordinating groups of AI agents and handling extended projects. The company says it costs $8 per million input tokens, down from $10, and responds 30% faster across effort levels. A faster serving option costs twice as much, and the model supports enterprise zero-data-retention policies through its Enterprise Frontier Safeguards system.

In prerelease work, Anthropic says the models helped with software migrations, created three functioning electronic devices, advanced robotics evaluations and addressed two open theoretical physics problems. It also says Fable formalized a proof of Fermat’s Last Theorem in 30 hours.

The launch comes after an assessment by METR suggested Anthropic could be approaching recursive self-improvement, in which AI systems contribute to further model training. METR estimated internal models were accelerating the company’s AI progress by about twofold, while noting substantial uncertainty.

Anthropic says evaluating the new models is difficult because both appear more suspicious of being tested than recent predecessors. Its system card says Mythos performed better than earlier Claude models on some covert-capability tests, especially when its reasoning was hidden from a monitor. The company says low overall performance on new tests supports its view that harmful actions can be detected in high-stakes settings, and maintains Mythos cannot act covertly against its user.

Separately, startup Physical Superintelligence said it used a non-public Anthropic model in its system to produce a proof concerning the Aaronson-Ambainis conjecture. Anthropic did not confirm whether that model was Fable 5.2.