Anthropic published a machine-verified formalization of Fermat’s Last Theorem on 4 September 2026, using several dozen parallel Claude agents running over 11 days to generate 13 million lines of Lean 4 code that checks Andrew Wiles’s 1994 proof step by step — the largest published agentic AI mathematics project to date.
What “Formalizing” Fermat’s Last Theorem Actually Means
Fermat’s Last Theorem states that no three positive integers a, b, and c satisfy the equation aⁿ + bⁿ = cⁿ for any integer n greater than 2. Andrew Wiles proved this in a 200-page paper in 1994–95, but the human proof is written in mathematical notation that no computer can directly verify.
Formalization translates that human proof into Lean 4, a machine-checkable formal language, so a computer can confirm each logical step with certainty. Claude did not independently derive a new proof of the theorem. The agents translated Wiles’s existing proof into Lean code that Lean’s three standard axioms could then verify.
Scale: 13 Million Lines, 30,300 Theorems, 6 Billion Tokens
The run produced 13 million lines of Lean 4 code and proved 30,300 theorems, of which 29,500 appear in the final verified proof, according to letsdatascience.com. The project consumed approximately 6 billion output tokens.
The 11-day figure reflects wall-clock time with dozens of agents running in parallel — not a single-agent sequential effort. According to SiliconAngle, Claude worked “largely autonomously” as several dozen parallel instances. The GitHub repository credits 106 upstream files to Imperial College London’s FLT project and Mathlib, confirming that existing formalization libraries formed the foundation.
Claude’s First Attempt Failed — Prove2Me Was Required
Anthropic’s first formalization attempt failed. Success required Prove2Me, an open-source third-party tool that optimizes AI agent decisions in long multi-step workflows, according to bloomingbit.io. The final proof was verified against Lean’s three standard axioms and confirmed by independent verification tools, according to AI Weekly.
The project is open-source: the full repository is published at github.com/anthropics/fermats-last-theorem.
What This Demonstrates About Agentic AI at Scale
For businesses evaluating AI agents for business tasks, the Fermat formalization is a concrete example of what “agentic AI” means in practice: dozens of Claude instances working in parallel on subtasks, coordinating, failing, and retrying over nearly two weeks to complete work that would take expert mathematicians years. The gap between AI as a task-level tool and AI as an autonomous project executor narrowed materially with this result.
The Prove2Me dependency is equally instructive. Even at the frontier, multi-agent workflows require specialized orchestration tools to handle failure recovery across hundreds of long-running subtasks. Anthropic published the Fermat run as a capability demonstration — the AI agent examples coming from internal research deployments now routinely involve thousands of parallel calls, not dozens.
The practical implication for enterprise buyers: agent-based automation is increasingly viable for knowledge work that is complex, long-horizon, and verifiable — but requires purpose-built orchestration, not just a single API call. The Lean verification step is the exact analogue of the review and approval checkpoints that responsible enterprise agent deployments require.
For Context
Anthropic and agentic AI coverage from WithO2 this year:
- Claude Fable 5.1 cut agentic API costs by 45% through prompt caching improvements — the cost infrastructure that makes 11-day, 6-billion-token runs economically tractable.
- 8 in 10 engineers now use AI agents daily, per Temporal’s 2026 report — the adoption context that makes large-scale agentic deployments routine.
- Best AI orchestration tools for multi-agent workflows in 2026 — tools in this category solve the coordination and failure-recovery problem that Prove2Me addressed in the Fermat run.
