Fermat Formalized in 11 Days: The Prove2Me Case

On a Tuesday in September 2026, Anthropic released a complete, computer-verified proof of Fermat's Last Theorem, produced by its Claude model.

Claude worked autonomously for eleven days, transforming the historic 129-page proof signed by Andrew Wiles into a version verifiable line-by-line by a machine, according to OfficeChai^[1]. Zero margin for error, the company claims.

A problem open for 350 years

In 1637, French mathematician Pierre de Fermat wrote in the margin of a book a statement as simple as it was persistent: no set of positive integers a, b, c satisfies the equation aⁿ + bⁿ = cⁿ for powers n greater than 2.

Fermat wrote that he possessed a "marvelous" proof, too long for the available space. That proof remains unfound to this day.

For over three centuries, the brightest mathematicians grappled with the enigma. In 1908, a monetary prize was established, equivalent to between one and two million dollars in today's currency, and in the first year alone more than 600 incorrect attempts were submitted.

Only in 1995 did British mathematician Andrew Wiles succeed in the feat, employing mathematical tools developed centuries after Fermat. His proof, 129 pages long, contained a critical error discovered during review: the correction required over a year of work with a collaborator, and the mathematical community spent months validating the entire framework.

The original idea: parallel agents on Prove2Me

Verifying a proof by hand remains slow and vulnerable to errors, even when experts dedicate months to checking. For this reason, the mathematical community has adopted tools called proof assistants, capable of validating mathematical logic automatically, such as Lean.

Translating a human proof into a language comprehensible to machines, a process called formalization, remains grueling work: every step the human considers obvious requires explicit explanation for the software.

For this reason, a group led by Kevin Buzzard at Imperial College London had launched in 2024 a multi-year project dedicated precisely to formalizing Wiles's proof. The estimate was years of coordinated human work.

Anthropic chose a different path. Through Prove2Me, a collaborative platform built by researcher Tianyi Peng and his team at Columbia University, dozens of Claude agents worked in parallel: they decomposed the proof into manageable fragments, proved each fragment, and rebuilt the whole. The technical detail is documented on Anthropic^[2]'s official research page.

Month one, verified results

The numbers released by Anthropic offer concrete scale of the work accomplished in eleven days.

• 13 million lines of Lean code generated by Claude
• approximately 30,300 minor theorems proved along the way
• a volume of code equal to over five times Mathlib, the primary formalized mathematical library of the community
• work accomplished in parallel by dozens of agents, coordinated via Prove2Me

Each figure comes from official company communications or from journalistic coverage cited above. At present, there remain no complete independent reviews published by third-party academic entities external to the project itself.

A correction of scale, not a reversal

The point of tension in this story comes from those already working on the same problem. Kevin Buzzard, who had led since 2024 the human formalization effort of Wiles's proof, commented publicly on what happened in a post dated September 4, 2026 on the Xena Project^[3] blog, acknowledging that Anthropic had beaten him to it.

This time gap, years of human planning against eleven days of machine work, remains the most instructive figure in the entire story. It also raises an open question about independent validation: a proof verified by a proof assistant still requires review by the mathematical community to gain full acceptance, exactly as happened with Wiles in 1995.

The willingness to acknowledge the achievement by the person leading the human project represents a sign of intellectual maturity, hardly classifiable as defeat.

What to take from this story

For a founder or CEO of a small business, the lesson concerns problem decomposition: dividing a complex task into independently verifiable fragments remains replicable with limited resources, even outside pure mathematics.

For a CTO or product manager, the relevant technical detail is the orchestration of parallel agents on a long and sequential task, coordinated through a shared platform like Prove2Me: an architectural pattern transferable to software verification workflows or quality control.

For a board or investor, the lesson concerns the speed with which computational capacity can compress a timeline estimated in years down to eleven days, shifting the bar of what appears achievable within a product cycle.

For a team manager, the case suggests a method: delegate to automated tools the tedious and repetitive parts of intellectual work, reserving for people the final review and critical evaluation.

An open question for every organization

Which problem, considered today too extensive to be addressed linearly, could benefit from the same decomposition into verifiable fragments that Claude applied to Fermat's proof?

The answer, for each organization, depends on the willingness to reorganize work before even choosing the technological tool.