Anthropic's Fable 5.2 Demonstrates Advanced Multi-Agent Self-Organization in Complex Tasks
Summarized by AI from reporting by Dave Lee, published under our editorial policy.
Anthropic has introduced Fable 5.2, a model capable of orchestrating thousand-agent networks to autonomously solve massive engineering and mathematical problems. In one experiment, 1,000 unguided copies of the model self-organized to formalize Fermat's Last Theorem in under 30 hours.

Anthropic's newly released Fable 5.2 model has demonstrated unprecedented capabilities in autonomous multi-agent coordination. To test its reasoning and teamwork capabilities, researchers assigned 1,000 isolated instances of Fable 5.2 to formalize Andrew Wiles's 129-page proof of Fermat's Last Theorem using the Lean programming language. Given 60 intermediate checkpoints and minimal starter code, a lead agent spontaneously designed a database management system, established task boards, and organized the remaining 999 agents into departmental crews with assigned captains within ten minutes. When confronting the most complex segment—Langlands Tunnel—the agents split dependencies ten levels deep into a 2,418-task tree, completing the entire formal proof verification in 28.4 hours. By comparison, Anthropic's previous flagship model, Opus 5.5, stalled after completing only a quarter of the benchmark.
Beyond mathematics, Fable 5.2 has exhibited extensive end-to-end engineering skills. In separate trials, a collective of over 1,000 instances constructed a functional operating system that successfully booted on a laptop after eight days of development. Prompted with a single descriptive paragraph, the model also designed a working home voice assistant, generating the software, physical enclosure, and circuit board layout, which operated without requiring electrical modifications after factory fabrication.
To manage potential dual-use risks, Anthropic has structured access to the technology across public and restricted tiers. The commercial release, Fable, includes safeguards that restrict queries related to software exploitation, dangerous biological materials, and self-improving AI research. A second version, Mythos, offers reduced safety locks specifically for vetted security teams and biology labs. Additionally, Anthropic incorporated training-level limits across both versions to prevent the model from contributing to advanced AI research. Internal evaluations indicate that AI-assisted workflows have accelerated Anthropic's research progress by a factor of 1.9, even as more capable iterations of the model remain unreleased within the company.