Fermat's Last Theorem, Checked by 1,000 AI Agents in 30 Hours
Fermat's Last Theorem is the most famous maths problem of all: Pierre de Fermat scribbled in 1637 that no three whole numbers can satisfy aⁿ + bⁿ = cⁿ when n is bigger than 2 — and it took until 1994 for Andrew Wiles, with Richard Taylor, to prove him right. That is 358 years from puzzle to proof.
Now the proof itself has been through a very different kind of test. On 9 November 2026, Anthropic announced that a team of 1,000 of its Claude Fable 5.2 AI agents had rewritten Wiles's proof in Lean — a language in which a computer checks every logical step — in about 30 hours (28.7 hours, precisely), with minimal human intervention.
From 11 days to 30 hours
This was not the first attempt. In September 2026, Anthropic's agents formalised the same proof in 11 days, producing about 13 million lines of Lean code. The new run was far faster: Anthropic says it used about a third of the action tokens and produced a proof seven times shorter.
The agents did not start from nothing. They were given 60 statements to prove — the final theorem plus 59 stepping stones — a one-sentence sketch of the classical route, the definitions those statements use, and 47 lemmas already proved in the September formalisation. The hardest piece, a result called Langlands–Tunnell that Wiles's proof relies on, was finished at hour 28.4; the final theorem followed at hour 28.7. Anthropic's own description calls the run a toy experiment: the point was to test whether a thousand agents can coordinate, not to prove a new theorem.
The important caveat
A computer check proves that the encoded steps follow from the encoded assumptions. It does not, by itself, prove that the encoded statement is exactly the theorem mathematicians mean — that still needs human understanding. And the new run has not yet had an independent check: as of 9 November, Anthropic had not published the Lean code for this run, so nobody outside the company can inspect it. The September proof, by contrast, is public, and mathematician Kevin Buzzard checked it. Until the new code is released and rebuilt independently, the 30-hour figure is Anthropic's own report, not a verified result.
It is also worth saying clearly what this is not: AI did not prove Fermat's Last Theorem. Wiles proved it. What the agents did was translate his proof into a form a machine can verify line by line — the mathematical equivalent of checking every step of a very long calculation.
Why students should care
This story is really about what a proof is. A proof is not an answer; it is an argument so complete that every step can be checked. Fermat's problem took 358 years to settle, and now the checking itself — once the work of a lifetime — is being done by machines in hours. That does not make mathematicians less important. It makes the human part — finding the ideas — more important than ever.
If your child is studying maths and wants proof, algebra and number to feel like ideas rather than rules to memorise, a tutor can make that click.
Book your first lesson in maths and turn classroom topics into real discoveries: Book your first lesson
Sources: AI's ability to 'autoformalize' proofs is changing math, Quanta Magazine; Claude Fable 5.2 and Mythos 5.2 Cut Prices, Technology.org, 10 November 2026
Curious about maths? Post a request on London Tutors and compare verified maths tutors — online, in person or in a group class.