Formalizing Fermat's Last Theorem in Lean
paperYour notes
The first end-to-end machine-checked proof of Fermat's Last Theorem. An internal Anthropic research model "roughly comparable to Claude Fable 5.1," running largely autonomously as dozens of parallel agents on a Claude Code-based multi-agent harness coordinated through the Prove2Me platform, wrote about 13 million lines of Lean and proved roughly 30,300 theorems (29,500 in the final dependency tree) in eleven days of wall-clock time ending August 18, 2026, consuming about six billion output tokens. The proof follows the Frey–Serre–Ribet–Wiles–Taylor–Wiles route as simplified in the Darmon–Diamond–Taylor exposition, builds on Kevin Buzzard's Imperial FLT project, flt-regular, and Mathlib, depends only on Lean's three standard axioms, and its statement was checked against Mathlib's own FermatLastTheorem with a comparator. Anthropic notes the proof is far longer than a human formalization would be, that about 7% of non-boilerplate lines are residue from failed attempts, and that early agent collaboration degraded before Prove2Me was adopted. Repository and 17-page report released September 4 (1,084 stars in a week).
November 2026 rerun with Claude Fable 5.2. At the Claude Fable 5.2 launch (November 9) Anthropic reported that a team of about 1,000 Fable 5.2 agents fully formalized the theorem in 30 hours with minimal human intervention, against eleven days for the first run. It started from minimal prompts and a list of intermediate theorem statements used to score progress, with no mathematical references, and the system card says it built on definitions and smaller results from the earlier formalization. Anthropic says the run used one-third as many action tokens and produced a proof seven times shorter. Claude Opus 5.5 in the same 1,000-agent setup completed only 26% of the formalization in 18 hours and made little further progress by 30. Both comparisons are single runs per model, and the figures are self-reported.