Anthropic Says Claude Formalized Fermat’s Last Theorem in Lean in 11 Days
On September 4, 2026, Anthropic published the first end-to-end computer-checked Fermat’s Last Theorem: ~13 million lines of Lean, ~29,500 theorems, ~6 billion output tokens, produced in about 11 days by agents using a model roughly comparable to Fable 5.1—not a new human proof.
TLDR
Anthropic on Friday, September 4, 2026 published “Formalizing Fermat’s Last Theorem.” Claim: first complete computer-checked FLT. Dozens of agents via Prove2Me (Tianyi Peng / Columbia). 11 days, ~6 billion output tokens, internal research model “roughly comparable to Fable 5.1.” **13 million** lines of Lean; ~29,500 intermediate theorems used (~30,300 proved); >5× Mathlib; Lean 4.33.1 + independent comparator. Follows Darmon–Diamond–Taylor exposition of Wiles—not a new human proof. Kevin Buzzard compiled it: “FLT: Anthropic has beaten me to it.” GitHub: anthropics/fermats-last-theorem. Tweet language “last month Claude completed” refers to the run; publication is September 4.
What this is (and is not)
| Item | anthropic.com/research Sep 4 |
|---|---|
| Result | Machine-checked formalization of known FLT |
| Not | A new number-theory proof or a Millennium Prize claim |
| Stack | Lean 4, Prove2Me, Fable-class agents |
Product-line de-dupe: not Fable 5.1 product (Sep 1), not OpenAI Navier–Stokes (Sep 8). Research publication.
Why this story matters
Autoformalization just ate a multi-year human Lean project in 11 days. That is a research-process story, not a party trick. Watch: whether Mathlib accepts the dump, and whether OpenAI’s NS claim four days later uses the same playbook.