Claude Proves Fermat's Last Theorem for the First Time with Tsinghua Talents

A proof milestone, but not a replacement for Wiles

Claude’s Fermat’s Last Theorem breakthrough arrived with a headline that is both spectacular and easy to misread. Anthropic announced on September 4, 2026, that Claude had produced what it called the first complete computer-checked proof of Fermat’s Last Theorem in Lean, working largely autonomously over 11 days [1]. The result does not mean that AI discovered Fermat’s Last Theorem from scratch, nor that Andrew Wiles’s 1995 proof has been superseded. It means that Claude helped convert a proof route for one of mathematics’ most famous theorems into a form that a proof assistant can verify step by step [1].

That distinction matters. Fermat’s Last Theorem states that no positive integers (a), (b), and (c) satisfy (a^n + b^n = c^n) for any integer (n > 2). Wiles, with Richard Taylor’s crucial contribution, established the theorem in the 1990s through deep machinery involving elliptic curves, modular forms, and Galois representations. Claude’s contribution is in formalization: translating and expanding mathematical reasoning into Lean code, so that a machine checks the logical dependencies down to foundational axioms [1].

The Tsinghua Yao Class connection

The human thread behind the project is central to the story. Anthropic identified Tianyi Peng as the researcher who set out to test whether Claude could make progress on formalizing Fermat’s Last Theorem [1]. The subject has drawn attention in China because Peng is associated with the Tsinghua University Yao Class talent pipeline, a highly selective computer science track founded around the academic vision of Turing Award winner Andrew Chi-Chih Yao. In that framing, the breakthrough is not only a Claude story, but also a story about how elite computer science training from Tsinghua’s Yao Class is now intersecting with AI, proof assistants, and high-level mathematics.

Peng’s role was not to manually type 13 million lines of Lean. According to Anthropic, human mathematical input was limited to occasional high-level instructions, while dozens of Claude agents handled the bulk of the work: defining concepts, proving intermediate theorems, and assembling those results into a path toward Fermat’s Last Theorem [1]. That combination of elite human direction and large-scale AI execution is why the result has been described as a breakthrough in computational proof methods.

What Claude actually built

The scale is the first striking number. Anthropic said Claude generated roughly 13 million lines of Lean code and proved 30,300 computer-verifiable theorems along the way, of which about 29,500 were used in the final proof [1]. A Chinese technology report summarized the same core figures, emphasizing that the project was larger than Lean’s main mathematical library by a factor of more than five [4].

The proof repository describes the artifact as a complete, machine-checked proof of Fermat’s Last Theorem in Lean 4, built on Mathlib and following the argument of Frey, Serre, Ribet, Wiles, and Taylor-Wiles [3]. Its formal statement says that for natural numbers (n), (a), (b), and (c), if (n \geq 3) and (a), (b), and (c) are positive, then (a^n + b^n \neq c^n) [3]. The repository also says its default build target checks that the theorem depends only on Lean’s standard axioms: propositional extensionality, classical choice, and quotient soundness [3].

That last point is important. In interactive theorem proving, it is easy to create the appearance of proof by leaving placeholders, adding unverified axioms, or relying on unsafe mechanisms. The repository states that no module contains sorry, added axiom, native_decide, unsafe, extern, implemented_by, partial def, or #eval, and that the proof was checked through both Lean’s kernel and an independent Rust-based Lean kernel called nanoda [3].

Prove2Me and the multi-agent workflow

The project did not succeed merely because Claude was asked to “prove Fermat.” Anthropic said early attempts failed because agents lost track of the project state and stopped coordinating effectively [1]. The successful run used Prove2Me, an open collaborative platform designed by Peng and collaborators to organize complex formalization work [1].

Prove2Me maintained a directed acyclic graph of theorem statements, helping agents decide what to prove next and how to reuse previous results [1]. It also separated theorem statements from proofs to accelerate compilation and reduce resource waste, and it maintained natural-language descriptions to help agents search for relevant intermediate results [1]. In practical terms, the system turned a huge mathematical mountain into thousands of smaller, checkable tasks.

Anthropic said the multi-agent effort consumed about six billion output tokens from an internal research model roughly comparable to Claude Fable 5.1 [1]. The Chinese report also highlighted this token scale, framing the project as a demonstration that formal mathematics can become a coordinated AI engineering workflow rather than a purely manual translation effort [4].

What Kevin Buzzard confirmed, and what he cautioned

Kevin Buzzard’s reaction is especially significant because he has led a major effort to formalize Fermat’s Last Theorem in Lean. On September 4, 2026, he wrote that Anthropic had beaten him to the finish line: an internal Anthropic model, using the Prove2Me platform, had formalized a complete proof of Fermat’s Last Theorem in Lean [2]. Buzzard said he had compiled the code base and run the comparator on it, and that “it checks out” [2].

His comments also add necessary restraint. Buzzard wrote that, mathematically, the work tells mathematicians “essentially nothing” new about Fermat’s Last Theorem, because the number theory community already accepted Wiles’s proof [2]. In his view, the importance lies elsewhere: the project shows what is now possible in autoformalization [2]. If thousands of pages of difficult literature can be formalized end to end by an AI swarm in 11 days, then future mathematics may be checked far closer to the moment it is produced [2].

Buzzard also noted that Anthropic’s proof follows the Darmon-Diamond-Taylor exposition of the Wiles-Taylor-Wiles argument rather than the modern proof route he has been formalizing himself [2]. He said his own project still matters because it aims to add reusable modern number theory to Mathlib and to build a dynamic document that humans can explore [2]. In other words, Claude’s achievement is a working formalization artifact; the longer-term community task is to make such formal mathematics elegant, reusable, and comprehensible.

Why this matters for AI and mathematics

The immediate impact is not that Fermat’s Last Theorem is now “more true.” The theorem was already proved. The impact is that a large language model, scaffolded by the right platform and overseen by expert humans, has performed a formalization task that experts expected to take years [1]. That changes the conversation about AI in mathematics from solving contest problems to maintaining, checking, and expanding the formal foundations of research.

Machine-checked formalization could reduce the burden of refereeing, especially as AI systems generate more mathematical claims [1]. Human-readable proofs will still matter: they explain ideas, reveal structure, and allow mathematicians to understand why a theorem is true. But formal proofs can serve a different purpose: they ensure that every logical step has been checked by a kernel that does not accept hand-waving [1].

The GitHub repository’s own warning is therefore apt. It calls the code a research artifact, not maintained and not accepting contributions [3]. It also says the Lean sources were produced by AI agents, are written to be checked rather than read, and that where a name and a statement disagree, the formal statement is what was proved [3]. This is not yet the polished textbook of the future. It is a massive, machine-checkable artifact showing that the future may arrive faster than expected.

The wider signal

For the Tsinghua Yao Class community, the Claude-Fermat result is a symbolic moment: one of its talents helped initiate a project that connects AI agents, proof assistants, and some of the deepest mathematical infrastructure of the twentieth century. For Anthropic, it is a showcase of Claude’s capacity to coordinate long-horizon technical work. For mathematicians, it is both an opportunity and a challenge.

The opportunity is a world where major proofs can be checked more quickly, libraries of formal mathematics can grow faster, and hidden assumptions in the literature become easier to expose. The challenge is that verification is not the same as understanding. Claude may have helped produce a proof that Lean accepts, but mathematicians will still need to decide how to curate, compress, explain, and integrate such artifacts into the living body of mathematics.

That is why the breakthrough should be read neither as hype nor as a curiosity. It is a proof-of-capability. Claude did not become Andrew Wiles. But with direction from researchers including Tsinghua Yao Class talent Tianyi Peng, it helped formalize Wiles’s world-changing theorem at a scale and speed that mark a new phase for computational proof.

Developments

1. Claude Proves Fermat's Last Theorem for the First Time with Tsinghua Talentseu.36kr.com · Sep 5, 2026, 1:40 AM UTC · 10/10

Sources from the last 72 hours

1. [1]Formalizing Fermat's Last TheoremSep 4, 2026, 12:00 AM UTC
2. [2]FLT: Anthropic has beaten me to itSep 4, 2026, 12:00 AM UTC
3. [3]GitHub - anthropics/fermats-last-theoremSep 4, 2026, 12:00 AM UTC
4. [4]Claude 11天自主证明费马大定理：1300万行Lean代码创史上最大形式化证明Sep 4, 2026, 4:00 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.