Claude Proves Fermat’s Last Theorem in Lean (2026)

A complete, sorry-free Lean formalization of FLT now exists in a public repository — verified not by Anthropic, but by Kevin Buzzard, the Imperial College mathematician leading the competing human effort who downloaded it, compiled it, and confirmed it holds.

What Happened

In new research published this week, Anthropic announced that Claude had produced a complete, machine-checkable Lean formalization of Fermat’s Last Theorem — a result documented in Anthropic’s research post and built openly on top of Mathlib, Imperial College’s own FLT-in-Lean project, and the existing flt-regular formalization. The critical word is formalization: Andrew Wiles proved Fermat’s Last Theorem in 1994. What Claude produced is a machine-checkable translation of that known result into the Lean proof assistant — no new mathematics, but an engineering feat of a different and measurable kind.

The single most important fact in the story is not in Anthropic’s post. Kevin Buzzard — the Imperial College mathematician who leads the multi-year community effort to formalize FLT in Lean — downloaded the public repository, compiled it himself, ran the kernel checker, and confirmed it holds. In a blog post he titled FLT: Anthropic has beaten me to it, Buzzard certified the proof is sorry-free (no unproven placeholders), depends on only the three standard axioms of Lean’s mathematics (propext, Classical.choice, Quot.sound), and that its statement matches Mathlib’s own definition of Fermat’s Last Theorem. He conceded the race. That is adversarial verification by the person with the most expertise and the most reason to find a flaw.

Two clarifications belong in any honest account. First, the p ≥ 17 restriction: the proof handles the small prime cases (p < 17) via the existing flt-regular formalization, which is legitimate — the smallest irregular prime is 37, so flt-regular covers all cases below that threshold. Anthropic disclosed this dependency; it is not a hidden gap. Second, Buzzard notes the formalization closes out Freek Wiedijk's 20-year "100 theorems" formalization list — a symbolic milestone for the formal mathematics community. What is not verified is Anthropic's account of how this was produced: the figures of roughly eleven days, approximately six billion output tokens, dozens of agents, a "Prove2Me" orchestration layer, and work done "largely autonomously" with occasional human nudges. Those numbers are Anthropic's own. The repository proves the output is valid Lean. It says nothing about how it was produced.

The Structural Read

The timing of this is not incidental. In the same week that a headline AI reasoning score moved from 98.6% to 62.7% under independent testing — as covered in FWMBA’s analysis of GPT-6 Astra’s independent benchmark collapse — Anthropic answered not with a new leaderboard entry, but with an artifact. The contrast is structural, not rhetorical.

Most AI capability numbers are self-reported against a setup the vendor controls. A Lean proof is the opposite kind of evidence. It either compiles under the kernel or it does not. The statement either matches the accepted definition of the theorem or it does not. There is no prompt, no harness, no framing that can fake a sorry-free proof. When the leader of the competing human project compiles your artifact and says “it checks out,” you have the strongest form of verification available in machine mathematics — external, adversarial, and binary. That is a genuinely different class of capability signal than a leaderboard number, and the distinction matters precisely because the week showed how fragile the alternative is.

But Buzzard’s actual critique — the sophisticated one, more interesting than any primacy dispute he already conceded — is worth reading carefully. He argues the formalization “tells us essentially nothing” mathematically new, made no contributions back into Mathlib, and produced no human-readable document explaining the proof. That is the correct framing of what this is: a verification and engineering feat, not a mathematical discovery. The distinction maps cleanly onto where agentic AI’s real, current value sits. Not inventing the new. Formalizing, checking, and completing the known — at a scale and speed that changes what is feasible to attempt.

The Astra-week contrast resolves into a capability-signaling playbook question. One approach: publish a score on a vendor-controlled setup and let the headline carry it until independent testing arrives. The other approach: publish an artifact in a public repository and let the rival compile it. Both are communication strategies. Only one is binary-verifiable on day one. That asymmetry — and what it suggests about which signals will matter as AI capability claims multiply — is the structural read this week deserves, connected to the broader through-lines tracked in FWMBA’s five through-lines synthesis.

Three Implications