Claude formalizes Fermat's Last Theorem in 11 days with 13M lines of Lean code

机器之心 · wechat · 2026-09-05

Anthropic says Claude completed a full Lean formalization of Fermat's Last Theorem in 11 days: 13 million lines of code, 30,300 machine-verifiable theorems (29,500 in the final build), with humans offering only occasional high-level hints. Mathematicians expected the effort to take years.

Context: Wiles' 129-page proof (1995) had never been fully formalized; Kevin Buzzard's 2024 community project was planned over years.

How: Claude followed the Darmon–Diamond–Taylor simplified route, with dozens of agents working in parallel on an internal model (Claude Fable 5.1), consuming 6B output tokens. The key scaffold was Prove2Me, an open collaboration platform that organizes theorems into a DAG, splits statements from proofs to cut compile time, and adds natural-language descriptions for retrieval — early runs without it collapsed, leaving 7% dead code.

Verification & limits: Lean checked the proof using only its three standard axioms, and statements match Mathlib; but the code is far longer and less elegant than human-maintained libraries.

Bonus experiment: three personal Claude Max accounts formalized Vinogradov's three-primes theorem in 3 days. Broader significance: AI proof generation may soon outpace human review, making machine-checked formalization new mathematical infrastructure.

Related event: Claude Produces First Formalized Proof of Fermat's Last Theorem in 11 Days(52 posts)→

Original post →

More from AGI Musings

AGI Musings channel →