Anthropic published the first complete computer-checked proof of Fermat's Last Theorem on September 4, 2026. Claude wrote it largely on its own across 11 days, producing about 13 million lines of Lean and proving 30,300 intermediate theorems, 29,500 of which survive in the final proof.
What Happened
The run followed a simplified version of Wiles's argument taken from Darmon, Diamond and Taylor. Human mathematical input was limited to occasional high-level instructions from Tianyi Peng, a researcher working with Anthropic and Columbia University. The model was a general-purpose internal research model that Anthropic describes as roughly comparable to Claude Fable 5.1, and the agents consumed about six billion output tokens getting there.
The artifact is public. Anthropic pushed the proof and a written walkthrough to the fermats-last-theorem repository, so the claim is inspectable rather than announced.
What This Enables
This is the largest documented long-horizon agent run that ends in a hard pass or fail signal, and the coordination layer is something you can open today. Prove2Me, built by Peng and collaborators at Columbia University, is the open platform that held the project together. It tracked which theorems depended on which, sped up compilation, and let each agent search and reuse work other agents had already finished. If you are building multi-agent systems that run for days rather than minutes, that shared dependency graph is the part worth copying.
Why It Matters for Creators
The failure modes are more useful than the headline. Anthropic says several early attempts collapsed because agents lost track of the project's state and stopped collaborating effectively, and that abandoned work still accounts for roughly 7 percent of the non-boilerplate lines in the finished proof. It also concedes the proof is likely much longer than it needs to be next to a human-curated library. Long agent runs rarely fail loudly. They drift, and durable shared state is what stops the drift. A separate human-led community effort to formalize the same theorem has been running for years at Imperial College London, which is the scale this compresses.
Key Details
Duration: 11 days of largely autonomous work.
Size: about 13 million lines of Lean, the largest Lean proof written.
Theorems: 30,300 proved, 29,500 in the final proof.
Compute: roughly six billion output tokens.
Model: an internal research model comparable to Claude Fable 5.1.
Waste: failed attempts left about 7 percent of the non-boilerplate lines.
What to Do Next
Clone the repository and read the walkthrough before the commentary, then study how Prove2Me modelled dependencies if you run agent swarms of your own. For a smaller open-weights take on the same problem, see our coverage of Mistral's Leanstral 1.5 proof agent.
