Claude Did in 11 Days What Mathematicians Expected to Take Years: the First Computer-Checked Proof of Fermat's Last Theorem, With Every Number Published

Anthropic has published something significant to both the AI and mathematics communities: an internal model produced the first end-to-end, computer-checked formalisation of Fermat’s Last Theorem in 11 days. Mathematicians had expected the work to take years.

We read Anthropic’s official research page directly. The numbers matter, and so does the question of how much a human did — because together they decide how this should be read.

1. The published numbers

2. What formalisation is, and why it matters

Checking a large mathematical proof by hand can take years. Formalisation converts the reasoning into a form a proof assistant such as Lean can verify mechanically, step by step.

So the achievement is not “AI proved Fermat’s Last Theorem” — Wiles proved it in the 1990s. It is that the proof now exists in a form a computer can check, and it took 11 days rather than years.

Anthropic also notes this was the final entry on Freek Wiedijk’s well-known list of 100 formalisation challenges, closing out a 20-year benchmark.

3. How much did a human do? Anthropic is straightforward about it

This is the part most easily overstated, so the official wording is worth quoting:

• The model worked “largely autonomously”, with occasional high-level instructions from researcher Tianyi Peng — for example, “Jacobian as a scheme sounds high priority”.
• Human mathematical input was “limited to occasional high-level instructions”.
• Mathematician Kevin Buzzard reviewed the result after completion.

So: not unsupervised, but the human input was direction rather than content. That distinction is the key one for anyone trying to gauge how long an autonomous run can now usefully be.

4. The enabling tool: Prove2Me

Anthropic credits the breakthrough to Prove2Me, described as “an open collaborative platform for formalizing mathematics” that maintained theorem dependencies, accelerated compilation, and enabled theorem search and reuse. Reporting adds that it maintains a directed acyclic graph of theorem statements and coordinates multiple Claude agents against it.

There is a transferable lesson for ordinary developers: what makes long autonomous work possible is usually not the model but the external state management and verification around it. Give an agent a structure that accumulates, is queryable, and validates automatically, and it stops getting lost several hundred steps in.

5. The weakness Anthropic states itself

The page says plainly that the proof is “likely much longer than it needs to be” compared with the concise standards of Mathlib, Lean’s mathematical library.

That is an honest and important line. 13 million lines is not a boast; it is volume that has not yet been refined. The everyday parallel is familiar: AI can finish the job, and finishing is not the same as finishing cleanly.

6. Still unverified

• Whether the internal model will ship as a product: Anthropic says only “roughly comparable to Claude Fable 5.1”. Unverified.
• The inference cost of those six billion output tokens: not published. Unverified.
• Prove2Me’s licensing and how usable it is for general developers: not detailed. Unverified.

Verified 2026-09-06. Sources: Anthropic research page, official PDF. Details may be updated. For how Claude’s plans differ, see our Claude tool page.