CONFIRMED: Claude proves Fermat's Last Theorem in Lean, with automatic verification

Anthropic disclosed in early September 2026 that its Claude model, orchestrated through the Prove2Me platform, produced the first complete, machine checked proof of Fermat's Last Theorem in the Lean 4 formal proof language. The official announcement came through an Anthropic research post and was confirmed by independent outlets, including SiliconANGLE, in a report dated September 4, 2026, and the AI Weekly newsletter. The theorem, stated by Pierre de Fermat in the 17th century and proven only in 1994 by Andrew Wiles after decades of attempts, now also exists as a formal proof, checked line by line by an automatic verifier, rather than only reviewed by human mathematicians.

Eleven days, dozens of agents, no human in the loop

The work was carried out largely autonomously by dozens of coordinated Claude instances over 11 days, consuming roughly six billion output tokens. Coordination happened inside Prove2Me, an open, collaborative math formalization platform built by Anthropic researcher Tianyi Peng together with collaborators from his research group at Columbia University. Prove2Me keeps a directed graph of theorem statements and distributes that graph across multiple Claude agents, which prove pieces of the problem in parallel and in sequence, reusing each other's results like a mathematical assembly line.

The scale is staggering: 13 million lines and 29,500 new theorems

The final result totals more than 13 million lines of Lean code and adds about 29,500 new theorems across major fields of mathematics, roughly five times the size of mathlib, Lean's main community math library. The most important point is not the size, it is the guarantee: Lean's proof checker verified the entire result from just three standard axioms of the language. That means the proof was machine verified, not simply written by an AI and taken on trust.

Why this is not just a math curiosity

Kevin Buzzard, a mathematician at Imperial College London and one of the most respected voices in the Lean formalization community, called the result an extraordinary achievement, reached in far less time than experts predicted. According to him, the feat could open the door to automatic formalization of a large part of the modern math literature, work that today depends on years of manual effort by specialists and that could, with support from AI agents supervised by verifiers such as Lean, move forward in a fraction of the time.

Why it matters for customer service and marketing automation professionals

For companies and agencies using AI in customer service, sales and marketing, the real value of this case is not the mathematics itself, it is the orchestration pattern: dozens of agents working in a coordinated way for 11 days, with no human in the loop, every step automatically checked by an unforgiving verifier before being accepted. It is a preview of what autonomous multi agent workflows can achieve when there is an automatic checking mechanism equivalent to Lean's. It is worth being honest about the limit of that comparison: Lean works because formal mathematics has a binary right or wrong and a verifier that accepts no middle ground. Business tasks, such as answering an upset customer, approving a refund or writing a campaign, are ambiguous by nature and do not have that kind of automatic referee. Anthropic's result is not evidence that autonomous agents can already run a company's critical processes alone, without supervision.

The direct takeaway: automatic verification, not blind faith in agents

The lesson for anyone designing automation is specific and should not be mistaken for a blanket endorsement of autonomous AI: the gain in scale and speed is only safe when it comes paired with a verification loop as rigorous as the problem allows, whether that is an automated test, an auditable business rule or human review at critical points. Multiplying agents without multiplying checks multiplies risk, not productivity. The responsible path for 2026 is to copy Lean's habit of verifying, not the illusion that orchestrating many agents is enough on its own.

Sources

Anthropic, Formalizing Fermat's Last Theorem: https://www.anthropic.com/research/formalizing-fermats-last-theorem | SiliconANGLE, Anthropic uses Claude to formalize proof of Fermat's Last Theorem: https://siliconangle.com/2026/09/04/anthropic-uses-claude-to-formalize-proof-of-fermats-last-theorem/ | AI Weekly, Anthropic's Claude Formalizes Fermat's Last Theorem in Lean: https://aiweekly.co/alerts/anthropics-claude-formalizes-fermats-last-theorem-in-lean