Andrew Wiles spent seven secretive years proving Fermat's Last Theorem, then still needed a fix for a hole in the proof a year later. Anthropic just had Claude turn the whole thing into fully computer-checked code in eleven days. Somewhere, a 17th-century margin note is feeling very called out.

Thirteen Million Lines of "Yes, This Is Actually True"

Anthropic announced that Claude produced the first complete, machine-verified formalization of Fermat's Last Theorem — converting the mathematical reasoning into Lean code that a computer proof assistant checks line by line, leaving zero room for the human error that can hide in a traditional peer-reviewed paper. The resulting proof runs 13 million lines of Lean, roughly five times the size of Mathlib, mathematics' largest existing formal library.

Dozens of Claude agents worked in parallel using an open-source coordination tool called Prove2Me, which tracked theorem dependencies as a graph so agents wouldn't lose the plot across a sprawling, months-spanning proof. Along the way they proved 30,300 theorems, 29,500 of which made it into the final chain, and burned through roughly six billion tokens of output. Mathematician Kevin Buzzard called it an "extraordinary autoformalization achievement" that proves the theorem "with no assumptions other than the axioms of mathematics."

Why Anyone Outside Academia Should Care

Formal verification has always been the tedious, unglamorous cousin of doing actual math — which is exactly why so little of the field's existing literature has ever been checked this rigorously. If AI can compress years of formalization labor into a couple weeks, it changes the economics of catching subtle errors in published proofs, not just producing flashy new ones.

The part that's easy to miss in the "AI does math" headline is the workflow, not the output: this was a swarm of agents coordinating on a shared, structured task over an extended run, largely unsupervised at the step level. That's the same pattern showing up in coding, research, and yes, business automation — long-horizon agentic work is quietly becoming table stakes.

Fermat's margin was too small to contain the proof; Claude needed 13 million lines instead — either way, the theorem's finally been checked all the way down.

James here — if you're curious what that same long-horizon agent pattern could automate inside your own business, let's talk: webtechkitchen.com/contact.

Source: Anthropic