Anthropic announced that Claude produced the first complete, computer-checked formalization of Fermat’s Last Theorem in the Lean proof assistant — about 11 days of largely autonomous multi-agent work, roughly 13 million lines of Lean, and about 29,500 intermediate theorems used in the final proof.
Unlike recent AI work that claims novel math, Anthropic frames this as verification: rewriting Andrew Wiles’s 1995-era proof (via a Darmon–Diamond–Taylor exposition) so a machine can check every step. The finished artifact uses only Lean’s three standard axioms and contains no unproved “sorry” placeholders. Kevin Buzzard reviewed the result and called it an extraordinary autoformalization achievement.
Early agent runs failed when collaborators lost track of project state. Success came after switching to Prove2Me, an open collaborative DAG platform designed by Anthropic researcher Tianyi Peng and Columbia collaborators, which let dozens of agents pick next theorems without stepping on each other. Anthropic says the campaign consumed about six billion output tokens from an internal research model roughly comparable to Claude Fable 5.1.
A companion experiment formalized Vinogradov’s Three Primes Theorem in three days using three consumer Claude Max plans collaborating through Prove2Me. The full FLT Lean proof is available on GitHub.
Sources: Anthropic Research; Kevin Buzzard / Xena; GIGAZINE
Leave a Reply