Anthropic details how 1,000 Fable 5.2 agents reproduced a formal proof of Fermat's Last Theorem

Anthropic says a team of 1,000 agents running an early version of Claude Fable 5.2 reproduced a Lean formalization of Fermat's Last Theorem in 28.7 hours, in what it calls a toy experiment. Its September formalization, by dozens of agents on an internal model roughly comparable to Fable 5.1 plus a researcher's direction, took 11 days and about 6 billion output tokens. Anthropic says the new run used one-third as many output tokens and produced a proof one-seventh the length, without the Prove2Me platform or a researcher steering it.

Setup. Each agent had a shell and file editor in its own lightweight virtual machine with no network access, Lean and Mathlib, its own Git branch with merge requests into main, and a shared Postgres database. Anthropic gave the team 60 statements to prove, a one-sentence sketch of the classical route, the needed definitions and 47 proved lemmas from the earlier work, with one lead and 999 helpers. It says it gave no other guidance on how to organize.

What the agents did. Anthropic says the lead built five database tables for orders, assignments, notes, subtasks and finished work in the first 10 minutes. Early on, helpers judged 47 of the 60 statements infeasible within the budget; the lead set them aside, then took them up again after the first 10 were proved in about 70 minutes. Around hour 18 it moved about 750 helpers onto Langlands–Tunnell, which its captain had put at under 1% odds of finishing in time. The helpers split it into a tree of 2,418 tasks up to 10 levels deep, and it was proved at hour 28.4, with the final theorem at hour 28.7.

Safety controls. Anthropic says that besides the sandboxes, a prompted classifier checked every shell command and file edit before it ran and a block would stop the whole team for human review, while an offline monitor reviewed the team's actions. It says none of the more than one million tool calls met the bar to be blocked. The run used only the action-based monitor; Anthropic added a monitor that also reads chain of thought and inter-agent communication shortly afterwards. It notes that it runs such experiments on models that may not have finished its pre-release alignment assessment.

Caveats. Anthropic says the result does not remove the need for domain-specific platforms like Prove2Me, and that agents tend to give up on hard problems early and misjudge their difficulty, so human encouragement can still help.