Fermat’s Last Theorem took Andrew Wiles seven years of secretive work to prove in 1994, closing a problem that had gone unsolved since a French mathematician scribbled it in a book margin around 1637. On September 5, 2026, Anthropic announced that dozens of instances of its Claude model had produced a complete, computer-checked version of that same proof in just 11 days. One leading mathematician called it extraordinary. Another, funded specifically to do this exact work, said it tells us nothing new about mathematics at all. Both are right, for different reasons.
What Just Happened: Claude Formalized Fermat’s Last Theorem
Over 11 days in August, dozens of instances of an internal Anthropic research model, roughly comparable in capability to Claude Fable 5.1, worked largely on their own to translate Wiles’s proof of Fermat’s Last Theorem into Lean, a programming language mathematicians use to write proofs that a computer can verify line by line rather than take on faith.
The agents produced 13 million lines of Lean code, more than five times the size of Mathlib, the community’s main library of formalized mathematics, and proved 30,300 intermediate theorems along the way, using 29,500 of them in the final proof. The whole effort consumed about 6 billion output tokens.
Mathematicians had expected this specific formalization task to take several years of dedicated human effort. Fermat’s Last Theorem was also the final unsolved entry on a well-known 20-year-old list of 100 major theorems mathematicians have been working to formalize one by one, meaning this result closes out that entire benchmark.
What Made This Possible: A Tool Called Prove2Me
Anthropic’s first attempt at this task failed. What changed the outcome was not a better model, but a platform called Prove2Me, built by Anthropic researcher Tianyi Peng with collaborators at Columbia University. Prove2Me keeps a map of every theorem statement the proof depends on, so dozens of Claude agents working at the same time can see what still needs proving and avoid duplicating each other’s work. It also splits theorem statements and their proofs into separate files to speed up compilation, and keeps a plain-language description of each statement so agents can find and reuse earlier work instead of starting from scratch.
Anthropic has been candid that the process was messy. Early agents made progress, then lost track of the project’s overall state and stopped collaborating effectively. Those abandoned attempts still ended up contributing around 7 percent of the lines in the final, successful proof.
Why Mathematicians Are Divided on What This Fermat’s Last Theorem Proof Means
Kevin Buzzard, an Imperial College London mathematician who has led a community effort to formalize the same theorem, reviewed Anthropic’s result and called it an extraordinary achievement, adding that the technique represents a genuine step toward automatically formalizing much of the modern mathematical literature, which he said could help catch errors buried in existing published proofs.
Not every mathematician shares that enthusiasm. One researcher who holds a multi-year grant specifically to formalize this same theorem pushed back hard, arguing the result tells us nothing new about mathematics itself. That criticism rests on an important distinction worth understanding on its own.
What Formalization Actually Means, in Plain English
Claude did not discover a new mathematical truth. Fermat’s Last Theorem has been proven and accepted by mathematicians since 1994. What Claude did was translate an already-accepted, human-written proof into a form so precise and mechanical that a computer program can check every single logical step, with no room for the small gaps in rigor that even careful human proofs sometimes contain. That is a genuinely different task from discovering new mathematics, and it is fair to say this result does not teach mathematicians anything new about Fermat’s Last Theorem itself.
Where the two camps actually agree is on what the achievement demonstrates about capability rather than about mathematics: a task that would have taken a skilled human formalization team years took a coordinated group of AI agents 11 days. Whether that is exciting or beside the point mostly depends on which question you’re asking.
Why This Matters Beyond Mathematicians
Formal verification, the broader technique this project relied on, is not just an academic exercise. The same kind of line-by-line, computer-checked rigor is used to certify that safety-critical software, like flight control systems or cryptographic code, behaves exactly as intended with no hidden logical gaps. If AI agents can now compress years of painstaking formalization work into days for a famous math theorem, similar techniques could eventually speed up rigorous verification work in other fields where mistakes are costly.
That said, it is worth applying the same healthy skepticism here that we’ve written about with other bold AI capability claims. Proving something already known, however impressively fast, is a different achievement than discovering something new, and it is worth keeping that distinction in mind the next time a headline blurs the two together. Our piece on the ‘No AI Fridays’ debate and what AI capability claims actually hold up to scrutiny covers a related pattern in how these stories tend to get reported.
For the full technical write-up, Anthropic’s own research post includes the complete timeline and excerpts of the agents’ work.
The Bottom Line on Fermat’s Last Theorem
Claude did not solve a mystery that stumped mathematicians for centuries. It took a proof humans had already worked out and translated it into a form no human could realistically check by hand, in a fraction of the time experts expected. That is a genuinely impressive feat of coordination and engineering, and it is also, quite reasonably, not the mathematical breakthrough some headlines are making it sound like. Both things are true at once, and that tension is the most honest way to read this story.
1 thought on “Fermat’s Last Theorem: The Incredible 11-Day AI Proof”