Anthropic Claude delivers first end-to-end computer-checked proof of Fermat's Last Theorem in 11 days
On Sep 4-5 Anthropic revealed that Claude, initiated by Tsinghua Yao Class alum Peng Tianyi, completed in 11 days the first end-to-end computer-verified formal proof of Fermat's Last Theorem, generating ~13M lines of Lean.
Between September 4 and 5 Anthropic announced that Claude had completed the first end-to-end computer-checked formal proof of Fermat's Last Theorem in 11 days, working largely autonomously. The project was initiated by Tsinghua Yao Class alumnus Peng Tianyi, using Anthropic Claude together with multi-agent collaboration and the Prove2Me platform. The model autonomously generated roughly 13 million lines of Lean code, proved about 29,500 intermediate lemmas, and the resulting corpus is more than five times the size of the Mathlib standard library, making it the largest Lean project to date.
Mathematician Kevin Buzzard reviewed the work and called it an extraordinary autoformalization achievement, noting that the proof is derived purely from standard mathematical axioms without additional assumptions. Andrew Wiles' 1995 classical proof remains the mathematical foundation, but Claude's work shows that AI can now undertake large-scale formal verification that previously took years, potentially accelerating the verification of new mathematical results and building trust in AI-assisted discoveries. Note that Claude did not propose a new mathematical proof route but instead rigorously transcribed Wiles' proof into Lean-verifiable form, omitting steps that Wiles had to fill in because of gaps in the original presentation, producing a chain of reasoning that is far more machine-checkable than the original paper.
The milestone demonstrates AI's practical engineering capability in large-scale formal verification. Combined with the parallel Enterprise Frontier Safeguards launch and the gated Mythos 5.1 release, Anthropic is simultaneously pushing model capability and safety boundaries to new highs.