Anthropic used an internal research model, comparable to Claude Fable 5.1, to formalize Andrew Wiles' 1995 proof of Fermat's Last Theorem. The process converted a 129-page mathematical proof into a computer-verifiable format using the Lean programming language.
The resulting file contains 13 million lines of Lean code, making it the largest file of its kind. To achieve this, the AI deployed several dozen agents that generated 6 billion tokens of output and proved 29,500 intermediate theorems.
The task was completed in 11 days, significantly faster than the several years mathematicians had expected. The breakthrough was made possible by integrating an open-source tool called Prove2Me, which optimizes the workflow for AI agents and reduces inference costs.
This achievement follows a recent advance where Anthropic used Claude to discover new information regarding the Riemann zeta function. Meanwhile, competitor OpenAI used its Astra model to solve Erdos problems and address questions in theoretical computer science.
