Anthropic releases its strongest pre-IPO models Fable and Mythos 5.2: a thousand AIs collaborate for 8 days to build a system, formalizing Fermat's Last Theorem in 30 hours

Beating AI News Flash: Anthropic releases Claude Fable 5.2 and Mythos 5.2. Both models are now capable of undertaking engineering and scientific research projects that previously required professional teams working over extended periods, including developing an operating system from scratch, formalizing mathematical theorem proofs, and solving long-standing unresolved problems in black hole theory.

Anthropic had Fable 5.2 develop an operating system starting from an empty folder. It orchestrated over 1,000 sub-agents, working continuously for 8 days, independently dividing labor, programming, testing, and reviewing code. The resulting system can boot on a laptop, featuring an independent kernel, disk and network drivers, a file system, graphical windows, and an installer.

In another experiment, approximately 1,000 Fable 5.2 agents spent 30 hours completing the formalization of an existing proof of Fermat's Last Theorem—that is, converting the mathematical proof into code that a computer can verify step by step. Previously, Anthropic took 11 days to complete a similar task using specialized tools and human researchers. Under the same configuration, Opus 5.5 ran for about 30 hours and completed only 26%.

Fable 5.2 also participated in solving two unresolved mysteries in black hole theory. One of them involved a complex calculation proposed in 2003 that had previously been difficult to complete.

Anthropic also had it independently design three electronic devices, including circuit boards, components, enclosures, and programs. After engineers at partner companies manufactured and assembled the devices according to the designs, all three devices functioned properly without any circuit modifications.

In official evaluations, Fable 5.2 achieved 49.1% on the FrontierCode 2.0 programming test, surpassing Opus 5.5's 45.5%. The robotics simulation test improved from 27.6% in the previous generation to 64.9%. Mythos 5.2 is oriented toward cybersecurity and biological research, achieving 89.6% in known vulnerability rediscovery tests, up from 78.3% in the previous generation.

Fable and Mythos use the same set of model weights, with the main difference being safety restrictions. Fable is open to all users, while Mythos has relaxed restrictions on cybersecurity and biological research and is currently available only to vetted institutions.

Anthropic has also proactively limited the model's training progress on tasks such as AI chip kernel optimization and model architecture improvement, keeping this capability roughly at the Opus 5.5 level to reduce the risk of AI accelerating the development of more powerful AI.

API pricing is set at $8 per million input tokens and $40 per million output tokens, a 20% reduction from the previous generation, with generation speed improved by up to 30%. Anthropic still recommends prioritizing Opus 5.5 for general programming and everyday knowledge tasks, while Fable and Mythos are primarily aimed at complex, long-running technical projects.