A Historic Milestone: Machine-Checking Andrew Wiles's Proof
On September 4–5, 2026, Anthropic announced a landmark achievement at the intersection of artificial intelligence and pure mathematics: an ensemble of Claude-powered autonomous agents completed the first complete, machine-checked Lean 4 formalization of Fermat's Last Theorem (FLT). The repository provides a fully verified computational formulation of the monumental modularity-lifting proof published by Sir Andrew Wiles in 1995.
Rather than inventing new mathematical principles, the Claude agents formalized the modern Darmon–Diamond–Taylor exposition of Wiles's proof, rigorously working through core foundational bridges including the Langlands–Tunnell theorem and Ribet's level-lowering theorem. Every logical step has been mechanically checked by Lean 4's type checker, demonstrating that the reasoning chains conform strictly to formal rules without heuristic hand-waving.
This computational breakthrough officially resolves the final remaining milestone on Freek Wiedijk's renowned list of 100 Famous Theorems to be formalized in a machine-verifiable proof assistant. For mathematicians and computer scientists, verifying FLT represents the culmination of decades of research into autoformalization and verified mathematics.
Scale and Architecture: 13.4 Million Lines of Axiomatic Proof
The scale of the generated artifact is unprecedented in the history of interactive theorem proving. The formalization spans approximately 13.4 million lines of Lean 4 code, condensing to roughly 10.5 million lines when excluding automatically generated boilerplate. Within this vast codebase, the agents proved 30,300 discrete lemmas and theorems, formally linking 29,511 of them into the primary directed acyclic dependency graph.
The computational campaign ran over approximately 11 consecutive days, starting on August 7, 2026. Dozens of parallel Claude agents generated roughly 6 billion output tokens, guided by Prove2Me, a directed acyclic graph agent coordinator developed by Tianyi Peng's research group at Columbia University. Prove2Me decomposed the overarching conjecture into modular, bite-sized formal subgoals that parallel workers could solve, verify, and stitch together iteratively.
Axiomatic integrity was maintained throughout the repository. The complete codebase compiles cleanly with zero unproven assumptions or `sorry` placeholders, relying exclusively on Lean's three foundational axioms. Independent external verifiers, including `comparator` and `nanoda`, inspected the generated proof terms and validated that the dependency chain holds end-to-end without logical circularity or system-level compromises.
Scaffolding and Attribution: Built on Human Foundations
While the engineering accomplishment is historic, critical nuances must be highlighted. Anthropic's system did not solve an open mathematical problem; Fermat's Last Theorem was definitively solved by Andrew Wiles over three decades ago. Claude's task was rigorous autoformalization—translating established human-written mathematics into strict, machine-executable code—rather than generating autonomous conceptual breakthroughs.
Furthermore, the project stands heavily on years of human labor. The proof directly imported and adapted 106 upstream files from the Imperial College London Fermat's Last Theorem initiative led by mathematician Kevin Buzzard, along with the community-maintained `flt-regular` repository and Mathlib foundations. Buzzard himself reviewed the repository, confirming its technical validity while detailing how prior human scaffolding made the run tractable.
Mathlib maintainers have noted that 13.4 million lines of machine-generated code presents clear structural drawbacks. The proof contains significant autoformalization verbosity and lacks the conceptual elegance or modular design required for standard upstream integration into Mathlib. Additionally, tech leaders clarified that this announcement is entirely separate from unsubstantiated online rumors suggesting Anthropic had solved the Navier–Stokes Millennium Prize problem.
Practitioner and Community Reaction
Within the formal verification and automated reasoning communities, the reaction has been largely celebratory. Specialists highlighted that successfully automating the final benchmark on Wiedijk's list proves frontier models can maintain multi-step logical consistency without catastrophic hallucination, provided they operate inside closed verification loops with determinative compilers.
However, experienced practitioners pushed back against initial marketing narratives that framed the run as an entirely unassisted AI breakthrough. Many noted that the system leaned heavily on Kevin Buzzard's modularization roadmaps and Tianyi Peng's Prove2Me orchestration architecture, arguing that without human-engineered mathematical scaffolding, parallel LLM swarms would have collapsed into search dead-ends.
Software engineers also voiced pragmatic skepticism regarding the practicality of million-line machine outputs. They emphasized that while Lean 4 mathematically validated every line, verifying and maintaining a machine-synthesized 13.4M line codebase poses severe cognitive friction for human teams, illustrating a growing tension between machine-verifiable correctness and human maintainability.
Strategic Implications for Business and Enterprise in Thailand
For enterprise technology executives and engineering leaders in Thailand, this milestone provides actionable validation that artificial intelligence is advancing beyond stochastic code generation into deterministic logic verification. The ability of LLM agent architectures to produce mathematically verified systems offers a preview of future enterprise quality assurance.
In Thailand's burgeoning digital banking, cross-border payment, and smart contract sectors, traditional testing methodologies often leave edge-case vulnerabilities unaddressed. High-assurance formal verification—historically cost-prohibitive due to the shortage of specialized verification engineers—can now realistically be augmented with coordinated AI agents to verify mission-critical protocols and API contracts.
Thai enterprises should look closely at hybrid architectures like Prove2Me, which combine agentic code synthesis with strict mathematical verification layers. Implementing machine-checked verification pipelines in enterprise CI/CD environments will enable local firms to harden financial platforms, critical infrastructure, and supply chain software against systemic failures with unprecedented rigor.
The milestone demonstrates that frontier AI can sustain verifiable mathematical reasoning across millions of lines of machine-checked code, unlocking automated formal verification for mission-critical banking, aviation, and software infrastructure.