AnswersPublished15 sources
How Claude Reportedly Formalized Fermat’s Last Theorem in Lean in 11 Days
Anthropic reports that Claude produced a complete Lean formalization of the already known proof of Fermat’s Last Theorem in 11 days: about 13 million lines of code and 29,500 intermediate theorems. The system reportedly split the proof into a dependency graph through Prove2Me, allowing agents to work on reusable def...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- Anthropic reports that Claude produced a complete Lean formalization of the already known proof of Fermat’s Last Theorem in 11 days: about 13 million lines of code and 29,500 intermediate theorems.
- The system reportedly split the proof into a dependency graph through Prove2Me, allowing agents to work on reusable definitions and lemmas in parallel while Lean checked completed proof terms.
- Lean’s checking establishes that the encoded theorem follows from its stated axioms and imports; it cannot by itself prove that every formal definition precisely captures its intended informal mathematical meaning.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How did Anthropic’s Claude reportedly produce the first end-to-end Lean computer-verified proof of Fermat’s Last Theorem in 11 days—includin. Article summary: Anthropic says Claude did not discover a new proof of Fermat’s Last Theorem (FLT); it translated a known modern route into a complete Lean artifact that Lean’s kernel can check. The reported advance is the scale and rela. Topic tags: general, education, general web, user generated, academic. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, water
Anthropic says Claude has produced the first complete, end-to-end computer-checked Lean formalization of Fermat’s Last Theorem (FLT). The important distinction is that Claude did not discover FLT’s proof: it converted an established modern proof route into a very large artifact that Lean can verify mechanically. Anthropic reports that this took 11 days of largely autonomous agent work and resulted in roughly 13 million lines of Lean code. 21
5
What Claude reportedly formalized
FLT says that, for natural numbers and an exponent of at least 3, the equation (a^n+b^n=c^n) has no nontrivial solution. In Mathlib’s Lean formulation, any solution must have (a=0), (b=0), or (c=0), which is equivalent to the familiar positive-integer statement. 31
The project followed the Darmon–Diamond–Taylor presentation of the Wiles/Taylor–Wiles strategy rather than inventing a new route through the mathematics. That choice matters: a formalization must encode not only the headline argument, but its definitions, hypotheses, supporting results, and all the connective reasoning that a conventional paper can leave implicit. Earlier work on the Imperial College FLT project also describes its route as a modern variant of the Wiles/Taylor–Wiles proof. 5
11
Why it took 13 million lines
A formal proof assistant does not accept “it follows by a standard argument.” Each inference must be expressed in a form the system can type-check, and each result must rest on formalized dependencies.
Anthropic reports approximately 30,300 computer-verified theorems produced during the project, with about 29,500 appearing in the final dependency closure. The final development spans fields including algebra, geometry, harmonic analysis, and number theory. 21
5
That scale helps explain why the 11-day figure should not be read as an 11-day replacement for the intellectual history of FLT. Anthropic also reports roughly six billion generated tokens. The effort included many candidate proofs, Lean checks, error-driven repairs, and the engineering needed to organize a long-running formalization. 5
The key technique: external coordination for many agents
The reported result depended on more than a single model maintaining a huge proof in context. Anthropic says earlier attempts made progress but struggled to retain shared project state. The successful run used a harness based on Prove2Me, a platform designed for collaborative mathematical formalization. 5
29
The central organizational idea is a directed dependency graph:
- Nodes represent outstanding definitions, lemmas, or theorems.
- Edges indicate which results are required by later results.
- Agents can select open, useful tasks, prove them in Lean, and make the resulting artifacts available to other agents.
- The system can then work upward from foundational components toward the final theorem.
Prove2Me describes this model as launching formalization “missions” to which agents contribute reusable formal proofs. 29 In practice, the graph converts one long, brittle task into many checkable subproblems, while creating durable project memory outside an individual model run.
What “computer-checked” means here
Anthropic says the final project contains no sorry placeholders—Lean’s mechanism for leaving a statement admitted rather than proved—and relies only on Lean’s three standard axioms. 5
This is stronger than an AI generating a convincing-looking proof in natural language. Lean’s kernel checks whether every encoded step follows under the system’s logical rules, axioms, and imported dependencies. Anthropic also says it verified that the terminal claim matches Mathlib’s established FermatLastTheorem formulation. 5
31
But kernel checking has a precise boundary. It verifies the formal statement that was encoded. Humans still need to assess whether the formal definitions and intermediate statements faithfully represent the intended informal mathematics. Independent compilation of the released artifact and expert review of its dependency chain are therefore essential tests of the reported achievement.
Why Kevin Buzzard called it extraordinary
Kevin Buzzard characterized the result as an “extraordinary autoformalization achievement,” according to Anthropic’s report. 21 The significance is not merely that agents can finish isolated Lean exercises. It is the claim that they assembled a deeply layered, reusable formal development at a scale previously expected to demand years of coordinated human labor.
If such work becomes reproducible and economical, formalization could help mathematics in several ways:
- More rigorous verification: formal systems can catch missing hypotheses, invalid transformations, and mismatched definitions that may be difficult to spot in prose.
- Reusable infrastructure: once a lemma is formalized, later projects can import it and rely on kernel checking rather than reconstructing every detail.
- A complement to peer review: machine-checkable proofs could reduce the burden of validating long technical arguments, while leaving mathematicians to evaluate meaning, relevance, and strategy.
These are prospective benefits, not automatic replacements for mathematical judgment. A formal proof can be logically valid while still formalizing the wrong statement, and the cost of producing a 13-million-line artifact remains substantial.
The right way to interpret the claim
The reported FLT result is a meaningful benchmark for AI-assisted formalization because the target theorem was already established and the final artifact is intended to be independently checkable. It should not be framed as Claude newly solving Fermat’s Last Theorem.
The strongest interpretation is narrower and more consequential: Anthropic reports that a coordinated agent system translated a major known proof into a Lean development at unprecedented scale, using graph-based task decomposition and formal verification as feedback. 21
5
29
The next questions are practical: whether outside researchers can reproduce the build, audit the formalization’s intended meaning, reuse its components, and achieve comparable results on other advanced mathematics. Those tests will determine whether this is a one-off engineering feat or a durable shift in how modern proofs are built and checked.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "How Claude Reportedly Formalized Fermat’s Last Theorem in Lean in 11 Days"?
Anthropic reports that Claude produced a complete Lean formalization of the already known proof of Fermat’s Last Theorem in 11 days: about 13 million lines of code and 29,500 intermediate theorems.
What are the key points to validate first?
Anthropic reports that Claude produced a complete Lean formalization of the already known proof of Fermat’s Last Theorem in 11 days: about 13 million lines of code and 29,500 intermediate theorems. The system reportedly split the proof into a dependency graph through Prove2Me, allowing agents to work on reusable definitions and lemmas in parallel while Lean checked completed proof terms.
What should I do next in practice?
Lean’s checking establishes that the encoded theorem follows from its stated axioms and imports; it cannot by itself prove that every formal definition precisely captures its intended informal mathematical meaning.