On September 4, 2026, Anthropic released a research result: Claude had translated an established proof of Fermat’s Last Theorem into Lean and completed machine checking over roughly 11 days. New coverage followed on September 5, but the research run had finished in August. The news is the public release of the result, not an AI suddenly solving an unsolved problem today.
Verification, not a new mathematical discovery
Fermat’s Last Theorem says that no three positive integers satisfy aⁿ + bⁿ = cⁿ for an integer exponent above two. Andrew Wiles’s proof was published in 1995. This project follows established mathematics and translates its reasoning into a formal language Lean can process.
A mathematical paper can leave out steps an expert considers obvious. A proof checker needs explicit definitions and deductions. The public repository documents build checks and a comparator: completion alone is insufficient; the axioms and the resulting statement must match the intended theorem in Mathlib.
How did the agents coordinate?
Anthropic reports that dozens of agents produced about 13 million lines of Lean. Prove2Me organized dependencies between theorem statements and supported searches for reusable results. That gave the agents shared project state rather than relying solely on an ever-growing conversation.
Imperial College London mathematician Kevin Buzzard wrote that he compiled the code and ran the comparator himself, successfully. He also distinguished a huge formal artifact from a readable, maintainable mathematical library others can reuse. Human understanding remains a separate task.
Not a one-click Claude subscription feature
The run used an internal research model roughly comparable to Claude Fable 5.1, a multi-agent harness and occasional high-level human directions. Anthropic disclosed about six billion output tokens, not an actual bill. It should not be presented as a free feature for ordinary accounts or assigned a research cost by multiplying public API prices.
The code is public, but the repository explicitly describes itself as an unmaintained research artifact that does not accept contributions. It is material to study and verify, not a supported product API or a promise that any problem pasted into a chat will receive a correct answer.
The important question: can the answer be checked?
For everyday AI users, the useful lesson is to separate generating an answer from checking it. Fluent explanations are not proof of correctness. Here, an explicit theorem, public code and formal verification tools make the result inspectable. The work demonstrates AI assistance on a large verification project; it does not mean every professional domain already has equally complete checks.
