A model formalizes Fermat's Last Theorem in Lean
On 4 September 2026 Anthropic reported a complete computer-checked formalization of Fermat's Last Theorem in the Lean proof assistant. Agents using an internal research model the lab compares roughly to Claude Fable 5.1, and a collaborative tool called Prove2Me, worked for eleven days, wrote about thirteen million lines of Lean, and used about 29,500 intermediate theorems in the final proof. The lab says human mathematical input was limited to occasional high-level steering, that Lean accepted the proof from its three standard axioms, and that the formalized statement matches a library statement of the theorem. The novelty they claim is the checking of a known proof, not a new mathematical argument.
The record ends here. Everything below this line is invented.
What if a proof counted as finished when only a machine had read it?
Part one
The Closed Margin
Leila refereed for a journal that had started to accept two objects where it used to accept one: a paper a person could read, and a formal proof a kernel could. The second object for the old theorem arrived on a Tuesday, thirteen million lines, and a one-line note that the root said proved.
She did not pretend to read it. She checked the statement against the library's own wording of the theorem, and she checked that the kernel was the ordinary one, three axioms, no extra. Both checks passed. The human essay that came with it was twelve pages and skipped, as essays do, the places the lines had swollen.
The editor wanted a decision before the weekend, because a second group had a claim in the queue and would build on this root if she passed it.
Leila opened the graph of dependencies at a node the essay never mentioned. The kernel called it proved. The comment above the node was a question a person had typed and no one had answered.
She set the decision to hold.
What’s real
- Agents formalized Fermat's Last Theorem in Lean in eleven days
- The proof is about thirteen million lines
- Lean checked it from standard axioms against a library statement
- The lab calls the novelty verification, not a new proof
What’s invented
- Leila's journal and its two-object rule
- A second group waiting to build on her pass
- An unanswered human question on a proved node
- Her hold on the decision
Part two: The Node With a Question
Part two isn’t on sale yet. Check back soon.
Filed under technology, formal mathematics, proof assistants, ai agents.