On September 4th, 2026 Anthropic published that their AI "Claude" created a complete and machine-verified formal proof for Fermat's Last Theorem: it consists of about 13 million lines of code, and about 30.000 verified theorems, created within about 11 days, built on Lean's library "mathlib" as well as preliminary work of the Buzzard community (in 2024, Kevin Buzzard started a project at Imperial College to formalize a complete proof in Lean, planned for many years and founded up to 2029). For more details see https://www.anthropic.com/research/formalizing-fermats-last-theorem.
Asking Claude (Opus 5) itself if it is possible to prove Fermat's last theorem also with Metamath, it answered: "Short version: in principle yes, in practice no. Foundations aren't the problem. set.mm is built on ZFC, and in the Grothendieck extension (ax-groth, Tarski's axiom) universes are available too - exactly what Wiles's proof appeared to need beyond ZFC. (Colin McLarty has shown in any case that this can be pushed down into far weaker systems.) So there is no logical obstacle, only a labor one. What has been proved so far is the case n = 4 (see flt4). ... Why it's nevertheless a size too big: Metamath has no tactics; every step stands there fully expanded. More serious is that set.mm simply lacks the necessary infrastructure: no schemes, no modular forms, no Galois representations, no class field theory. That is the actual proof - the number theory of FLT is only the last line. Mathlib had a large part of it after years of community work, and even building on that, the Lean formalization comprised some 30,000 additional theorems; set.mm in total holds only an order of magnitude more than that. ... So it would be feasible - but it would be a project on the same scale as everything assembled in set.mm since 1993."
Nevertheless, this and the previous subsections provide some starting points for a formal proof by Metamath: According to the blueprint provided by Kevin Buzzard, see https://imperialcollegelondon.github.io/FLT/blueprint/, Lemma 2.1 is already proven (see fltoprm). The next step would be to prove Lemma 2.2 (FLT for n = 3, see flt3 above). Corollary 2.3 follows immediately from Lemmas 2.1 and 2.2 (see fltoprmgt3).
Finally, asking Claude if it makes sense to translate the Lean proof into the Metamath language, it says: "Possible in principle - sensible, not really. Foundationally there's no obstacle. Mario Carneiro showed that Lean's type theory is consistent relative to ZFC with n inaccessible cardinals for all n < . set.mm, via Tarski's axiom (ax-groth), supplies arbitrarily large inaccessibles - so the required model is within reach. And there's precedent: in 2014/16 Carneiro translated proofs from the OpenTheory format (HOL Light, HOL4, Isabelle) into Metamath, first into a HOL formalization hol.mm, then embedded into set.mm. The exercise has been done once already. But Lean is a different league from HOL. Three problems: * The translation isn't a rewrite. ... * The transfer problem. ... * The scale. Lean's elaborator performs an enormous amount of invisible work: instance resolution, implicit arguments, universe unification, definitional unfolding. Metamath has none of it. 13 million lines of Lean on top of over a million lines of mathlib plausibly expand into the billions of steps - against the few tens of thousands of statements that make up all of set.mm. The real objection, though, isn't feasibility but payoff. ... Nothing for set.mm. The database is curated and maintained for readability; a generated dump would contradict its purpose and wouldn't be admitted into the main database anyway. What does transfer is the blueprint. The lemma decomposition, the definitional choices, the proof strategy - that's language-independent, and it's precisely the part that took years of work to produce. Buzzard's blueprint for the FLT project is literally that document. A native Metamath development following it would be the sensible path, and it would reasonably start at n = 3.".
The last sentence is actually the way we want to continue (as described above)...