What did the research establish?

AI helped turn a major human proof into a complete formal object that can be checked by software. The accomplishment is in formalization; the original mathematical credit remains with the people whose work it encodes.

01

What changed

Fermat’s Last Theorem says that no positive integers a, b and c satisfy aⁿ + bⁿ = cⁿ when the whole-number exponent n is greater than two. Andrew Wiles, with Richard Taylor’s help repairing the argument, established the result in the 1990s. Turning that reasoning into a complete machine-checkable proof remained a separate challenge.

On September 4, 2026, Anthropic announced that Claude agents had completed that translation into Lean. Anthropic reports 11 days of largely autonomous work. The agents followed an exposition of the Wiles–Taylor–Wiles argument by Darmon, Diamond and Taylor; they did not discover the theorem anew.

Kevin Buzzard, who leads a separate Fermat formalization project at Imperial College London, reports compiling the code and running comparator, a tool that checks the proved statement against a reference statement. His check passed. The released FinalCheck file also connects the result to Mathlib’s statement of Fermat’s Last Theorem and checks the axioms used.

02

What this could change for you

A written proof leaves many steps implicit for expert readers. Formalization makes those steps precise enough for a proof checker. This result demonstrates that AI can help assemble a large, interconnected formal proof. It does not show that every AI-generated mathematical argument is correct.

Humans supplied the mathematical proof, Lean, Mathlib and earlier formalization work. The repository credits reused or adapted code from the Imperial project and flt-regular. Researchers built Prove2Me to coordinate the agents’ proof dependencies, and Tianyi Peng provided occasional high-level direction. Claude produced and assembled further formal code within that infrastructure.

The proof and reading materials are publicly available on GitHub. This is a research artifact, not a one-click checking service. The system used an internal research model; buying an ordinary chatbot subscription is not a documented way to reproduce this project.

From research to everyday use

Can you use this yet?

This is a research result rather than a consumer tool. Practical use depends on turning the method into a tested application; the brief does not establish when that will happen.

Check the study’s limits
03

What it does not prove

The guarantee concerns the exact formal statement and its assumptions, with trust in the checking software. The final result uses Lean’s standard axioms. A checked proof does not automatically make every intermediate label or generated English explanation accurate, and it does not substitute for an understandable mathematical account.

Buzzard emphasizes that this adds no new mathematical result to the existing proof. His separate project still has goals involving reusable library contributions and human-readable exposition. The released AI artifact is large and marked unmaintained. AI Solved X checked the public sources and expert assessment; we did not independently compile the full proof.

Primary research

Formalizing Fermat’s Last Theorem

Anthropic research announcement · September 4, 2026

View the research ↗