1,000 Fable 5.2 Agents Fully Formalize Fermat's Last Theorem in 30 Hours

Anthropic published results from an AI4Science experiment: roughly 1,000 agents running an early version of Claude Fable 5.2, with no Prove2Me orchestration platform and no researcher guidance, collaborated using only two simple primitives—a shared database and git—to fully formalize the proof of Fermat's Last Theorem into a computer-verifiable Lean version within 30 hours. This means translating every step of Andrew Wiles's 1995 proof into a form machines can verify step by step, a long-standing goal of the mathematics community.

Confirmed

• The thousand-agent Fable 5.2 team reached 100% completion in 30 hours, and 86% in 18 hours (relayed by the Claude official account, @minchoi, and @likeastar20)
• Under the same setup, Opus 5.5 completed only about 26%-27% in 18 hours
• The collaboration mechanism was minimal: no orchestration platform, no researcher guidance, just a shared database and git (@RLanceMartin)
• The team's kernel-verified proof is only about one-seventh the length of the original (@RLanceMartin)

Why it matters

• The experiment validates the ability of large-scale multi-agent systems to perform high-level mathematical formalization without centralized orchestration, a landmark case for AI-assisted scientific research
• The stark gap between Fable 5.2 and Opus 5.5 (86% vs. roughly 26%) shows how model capability differences are amplified in long-horizon collaborative tasks
• A shared database plus git sufficed to support thousand-scale agent collaboration, suggesting complex orchestration frameworks are not strictly necessary—a useful reference for multi-agent system design

2026-11-10 ~ 2026-11-10 · 12 related posts

• [source] 1,000 Fable 5.2 agents formalize Fermat's Last Theorem in 30 hours; Opus 5.5 stalls at 26% — claudeai · 2026-11-10
• 1,000 Fable 5.2 agents fully formalize Fermat's Last Theorem in 30 hours — likeastar20 · 2026-11-10
• [source] Anthropic's 1,000-agent Fable 5.2 team reproduces a formal proof of Fermat's Last Theorem — RLanceMartin · 2026-11-10
• 1,000-agent team formalizes Fermat's Last Theorem proof under safety boundaries — RLanceMartin · 2026-11-10
• With just a shared database and git, 1,000 agents wrote a kernel-checked proof 7x shorter — RLanceMartin · 2026-11-10
• Anatomy of a 1,000-agent run: sandboxing, blocking classifiers, and offline monitoring — RLanceMartin · 2026-11-10
• Anthropic's 1,000 agents built their own database tables and split into subteams to work in parallel — RLanceMartin · 2026-11-10
• [source] Hardest Fermat lemma spawned a 2,418-task tree across ~750 agents, 10 levels deep — RLanceMartin · 2026-11-10
• 750 agents coordinate via a 2,418-task tree to prove a hard theorem — RLanceMartin · 2026-11-10
• Scaling a team of 1,000 Fable 5.2 agents safely yields a formalized proof of Fermat's Last Theorem — EricBuess · 2026-11-10
• New benchmark: formalizing Fermat's Last Theorem — Fable 5.2 hits 100%, Opus 5.5 stalls at 27% — appenz · 2026-11-10

1 near-duplicate retellings: minchoi