Anthropic said Claude produced the first complete, computer-checked formalization of Fermat’s Last Theorem in the Lean proof assistant, working largely autonomously over 11 days, according to a Sep. 4 company research post. The agents wrote about 13 million lines of Lean and proved 29,500 of roughly 30,300 attempted intermediate theorems, drawing on about 6 billion output tokens from an internal research model Anthropic described as roughly comparable to Claude Fable 5.1.
The result is verification, not a new proof of the theorem. Anthropic said Claude followed a simplified Darmon–Diamond–Taylor route to Andrew Wiles’s 1995 proof, and that Lean checked the finished artifact against Leans three standard axioms. A comparator confirmed the theorem statement matches Mathlib’s own FLT statement. Failed early attempts still contributed about 7% of the non-boilerplate lines in the final proof, the post said.
Anthropic shared the proof with Kevin Buzzard of Imperial College London, who has led a parallel EPSRC-funded community formalization of FLT since 2024 on a multi-year timeline. Buzzard called it “an extraordinary autoformalization achievement” that “proves Fermat’s Last Theorem with no assumptions other than the axioms of mathematics,” and said the artifact is robust enough to build on.
The campaign used Prove2Me, an open collaborative formalization platform from Anthropic researcher Tianyi Peng and collaborators at Columbia University, to keep a DAG of theorem goals, split statements from proofs, and let dozens of agents work in parallel without losing project state. Anthropic published the Lean proof on GitHub with a written walk-through.
GIGAZINE and other secondary outlets amplified the research post on Sep. 7. Anthropic framed the work as a step toward autoformalizing large parts of the modern literature so AI-generated math can be checked without years of human review — not as a claim that Claude discovered Wiles’s mathematics from scratch.