Anthropic has published a complete, machine-checked formalization of Fermat's Last Theorem in the Lean proof assistant, releasing the codebase as an unmaintained open-source research artifact. In its September 4 announcement, Anthropic stated that the formalization follows an exposition of Andrew Wiles's known proof by Henri Darmon, Fred Diamond, and Richard Taylor. The project focuses on automated verification of existing mathematics rather than proving an unsolved conjecture.
According to Anthropic, Claude produced the formalization largely autonomously over 11 days using dozens of agents directed by occasional high-level human guidance. The company reports that the effort consumed roughly six billion output tokens and generated 13 million lines of Lean code, driven by an internal general-purpose research model roughly comparable to Claude Fable 5.1.
After initial multi-agent coordination failed, the team integrated Prove2Me, an open collaborative platform described in an August paper by Shuze Chen and colleagues that enables AI agents to contribute machine-checkable proofs and reuse verified results through specialized collaboration harnesses.
Inspecting and Verifying the Codebase
The code is available under the Apache 2.0 license in Anthropic's GitHub repository, pinned to Lean 4.33.1 and Mathlib v4.33.0. The codebase builds upon prior mathematical foundations from the Mathlib community, Imperial College London's FLT project, and flt-regular.
Practitioners inspecting the work do not need to rerun the multi-agent generation system. The repository includes offline HTML proof navigation, dependency graphs, and exact formal statements, allowing researchers to trace definitions and proof paths without running a complete rebuild. The formal statement itself serves as the authoritative definition, as theorem names alone do not guarantee intended mathematical semantics. The repository's final check guard requires the derivation to rely solely on Lean's three standard axioms, without placeholder proofs or auxiliary axioms, matching Mathlib's formal theorem statement.
Rebuilding and verifying the artifact locally remains resource-intensive. While maintainers report clean rebuilds and comparator replays confirming compatibility with Mathlib, reproducing the verification requires high-memory infrastructure, with maintainers reporting a 230 GB peak for comparator replay and recommending 300 GB of RAM. Anthropic released the repository as a static snapshot and is not accepting contributions or maintaining the codebase.