A Historic Formalization: 13 Million Lines of Machine-Checked Lean 4
Anthropic publicly released a landmark verification artifact on September 4–5, 2026: a complete, machine-checked Lean 4 formalization of Fermat's Last Theorem (FLT). Marking the largest formal proof ever compiled in the Lean language, the codebase encompasses roughly 13 million lines of code—approximately 10.5 million lines when excluding autogenerated boilerplate—distributed across 60,475 modules.
The scale of the proof is unprecedented in formal mathematics: it formalizes and proves 29,511 intermediate theorems alongside roughly 533,000 local lemmas. The computational run was initiated in the early hours of August 7, 2026, and achieved full verification in approximately 11 days, proving the final theorem on August 17, 2026, before open-source release this September.
Rather than attempting novel mathematical discovery, the effort represents a full autoformalization of the 1995 Darmon–Diamond–Taylor exposition of the Frey–Ribet–Wiles–Taylor proof. The system directly built upon the Lean ecosystem's foundational library, pinned to Lean 4.33.1 and Mathlib v4.33.0, and adapted 106 existing files from the Imperial College London FLT project led by Kevin Buzzard alongside the open `flt-regular` repository.
Verification Soundness and Multi-Agent Orchestration
Crucially, the verification is mathematically sound and devoid of synthetic shortcuts. The standard Lean kernel verified the entire proof relying strictly on Lean's three foundational axioms: `propext`, `Classical.choice`, and `Quot.sound`. The codebase contains zero placeholder `sorry` statements, introduces no unproven external axioms, and uses no `native_decide` escapes.
To provide independent assurance against implementation bugs in Lean itself, the external `nanoda` kernel (version 0.4.13, written in Rust) was deployed to parse the compiled proof. It validated and accepted all 1,052,234 declarations without error, successfully meeting the verification standard established by the `leanprover/comparator` challenge.
The mechanical engine behind this execution involved dozens of coordinating Claude agent instances communicating through Prove2Me, an open collaborative theorem-proving harness developed by Anthropic researcher Tianyi Peng's team at Columbia University. The entire 11-day multi-agent workflow consumed roughly 6 billion output tokens generated by an internal Anthropic research model comparable to Claude Fable 5.1.
Practitioner Reactions: Landmark Breakthrough vs. Tactical Bloat
Within the formal mathematics and verification communities, the release was widely hailed as a watershed achievement, resolving the final remaining milestone on Freek Wiedijk's famous list of 100 formalization challenges. Formal verification specialists embraced the fully checked codebase, arguing that securing a computer-certified end-to-end proof of FLT outweighs concerns about aesthetic inelegance.
Conversely, library maintainers and software engineers raised immediate concerns regarding downstream code usability and maintainability. Commentators noted that the generated Lean source is excessively verbose, machine-written, and non-idiomatic, suffering from severe 'tactical bloat.' Because these proofs are difficult for humans to read, refactor, or maintain, there is no verified commitment that Mathlib will directly merge or incorporate any meaningful portion of the 29,511 intermediate lemmas.
Furthermore, practitioners stressed the boundary between verification and discovery: the system successfully automated the formalization of established human mathematics, but did not invent new mathematical theorems. Meanwhile, speculative rumors circulating among community members claiming Anthropic had quietly solved the Navier–Stokes Millennium Prize problem remain completely unconfirmed.
Implications for Enterprise and Critical Software in Thailand
For enterprise executives and technology leaders in Thailand, Anthropic's achievement illustrates an essential paradigm shift: AI reasoning is evolving from probabilistic generation—plagued by hallucinations—into deterministic, mathematically verifiable execution when coupled with external proof checkers.
In Thailand’s burgeoning digital economy, industries governed by zero-tolerance reliability standards—such as financial institutions developing core banking or national payment gateways, mission-critical energy utilities, and smart contract platforms—can leverage automated formal verification. Multi-agent theorem proving provides a credible roadmap for certifying safety-critical firmware and eliminating high-impact logical vulnerabilities prior to deployment.
Nonetheless, the operational takeaway for Thai engineering organizations is the trade-off between volume and maintainability. Generating millions of lines of machine-derived logic introduces massive technical debt if not structured for human comprehension. Enterprises adopting automated agents must design verification guardrails that balance mechanical soundness with operational clarity, ensuring critical systems remain inspectable and manageable over their entire lifecycle.
This milestone marks a decisive shift in AI reasoning, demonstrating that automated agents can mathematically verify extreme-scale logic with zero hallucinations—laying new ground for enterprise mission-critical software verification.