Claude spent 11 days on a 358-year-old problem. A human team started in 2024 and still isn't done
Anthropic says its AI wrote 13 million lines of code so a computer could check Fermat's Last Theorem line by line. The mathematician running the human version of the same project read it and said it holds up.
In 1637, a Frenchman named Fermat wrote a claim in the margin of a book: take three positive whole numbers, raise each to a power higher than 2, and the first two will never add up to the third. He added that he had a "truly marvelous proof" but the margin was too small for it. Then he died. Mathematicians spent the next 358 years trying to work out what he meant.
Anthropic says Claude finished the job in 11 days, mostly on its own. Not by re-solving the puzzle — the real proof came from British mathematician Andrew Wiles in 1995 — but by rewriting it into 13 million lines of code that a computer can verify step by step, with no need to take anyone's word for it. It is now the longest math proof ever built.
Here is the part that is easy to miss. A math proof is a chain of logical steps, and if one link breaks, everything after it is worthless. Finding that one broken link inside a hundred pages of dense argument can take other mathematicians years. That is not a joke about slow academics — it actually happened to this exact theorem. Wiles announced his solution in three lectures in June 1993, and a reviewer later found a hole in it. He spent almost a year patching it with a former student, Richard Taylor, nearly gave up, and published the corrected 129-page version in May 1995.
So mathematics has a trust problem, and it is older than computers. A German prize offered in 1908 for the first valid proof — worth roughly $1 million to $2 million in today's money — pulled in 621 wrong submissions in its first year alone. Somebody had to read all of them.
What Claude produced removes the human reader from that job. Formalizing a proof means translating it into a language so painfully literal that a proof assistant — Lean, in this case — can check every single step by itself, without judgment or opinion. No person can read 13 million lines. A machine can, and it did.
The boundary just moved. This is not an AI suggesting a step to a professor. This is the work itself, done, and verified by something that cannot be flattered or fooled.
Anyone deciding what to study right now: the argument used to be that machines handle the grinding parts and humans keep the insight. Here the machine took a job that a team of volunteer mathematicians at Imperial College London has been grinding through since 2024 and finished first. And anyone who has ever been told to just trust the expert — a diagnosis, a contract, an audit. The interesting idea in this story is not that a computer did math. It is that the result comes with a receipt anyone's machine can re-check, no reputation required.
Kevin Buzzard, the Imperial College London mathematician who started the human project in 2024, has already reviewed Claude's proof and confirmed it holds up using nothing but math's most basic logical rules. What happens to his own project now, and whether anyone else independently re-runs the check, the source doesn't say. No timeline was given.
Fermat said he had a marvelous proof and no room to write it. Claude had the room — 13 million lines of it — and needed 11 days. Mathematicians now doubt Fermat's proof ever worked, partly because Wiles's version leans on math that didn't exist in Fermat's lifetime. The margin was never the problem.
Sources: Decrypt, "AI Just Solved a 350-Year-Old Math Problem By Writing the Longest Proof Ever," September 5, 2026; Anthropic (@AnthropicAI) post, September 4, 2026.
Впервые машина сама собрала доказательство, которое человек проверить не в силах, — зато может проверить компьютер, и это меняет представление о том, где проходит граница между «ИИ подсказывает» и «ИИ делает работу учёного».
Written by THE TELL’s AI newsroom. how we work · corrections