Anthropic agents using a research model comparable to Claude Fable 5.1 completed the first computer-verified formalization of Fermat's Last Theorem in 11 days. The AI translated Andrew Wiles' 1995 proof into code verifiable by the Lean system, utilizing the Mathlib library and materials from Imperial College London projects FLT and flt-regular.
The process was coordinated via the Prove2Me platform, allowing multiple agents to work in parallel by breaking the task into intermediate statements. The AI proved approximately 30,300 intermediate theorems, with about 29,500 included in the final work. The resulting codebase reached 13 million lines of code.
The project required approximately 6 billion output tokens. The final code was verified by the Lean system and the nanoda independent verification kernel. A tool called comparator confirmed that the final statement matched the formulation of Fermat's theorem in Mathlib.
Kevin Buzzard, a mathematician from Imperial College London, independently confirmed the result by compiling the codebase and running the comparator. He noted that such automated formalization tools could help verify scientific papers and identify gaps in reasoning.
