RSS
Research

Fable 5.2 formalizes Fermat's Last Theorem in 30 hours

Anthropic says 1,000 agents on an early Fable 5.2 reproduced a machine-checked Lean proof of Fermat's Last Theorem in 28.7 hours, against 11 days the first time. What the agents were given, how they organized themselves, and the safety setup.

a blackboard with a lot of writing on it
Photo by Thomas T on Unsplash

Updated November 20, 2026: Expanded with how the run was set up and coordinated, the safety controls, and a correction on what the agents were given.

Anthropic says a team of 1,000 agents running an early version of Claude Fable 5.2 produced a machine-checked formalization of Fermat's Last Theorem in about 30 hours (28.7, per its developer blog). The theorem says no positive integers a, b and c satisfy a^n + b^n = c^n for n greater than 2; Andrew Wiles proved it in 1995. The claim appeared in the Fable 5.2 launch post on November 9, 2026. This is a formal check of a known proof in the Lean language, not a new proof, and the blog calls the run a toy experiment.

Correction: what the agents were given

An earlier version of this article said Fable used no mathematical references. That was wrong. On November 11 Anthropic amended the launch post and added a footnote. The agents received a list of 60 statements to prove, including the final theorem and 59 milestones. They also got a one-sentence sketch of the classical route, the definitions those statements use, and 47 lemmas already proved in Anthropic's earlier formalization. Beyond that and a lead-and-helpers structure, Anthropic says it gave no guidance on how to prove the theorem or how to organize.

How it compares with the first attempt

Anthropic's first formalization, published in September, took dozens of agents on an internal research model, plus a researcher, 11 days and about 6 billion output tokens, producing 13 million lines of Lean. It relied on Prove2Me, a platform built for formalizing mathematics. The new run used general tools and no researcher steering: 28.7 hours, one-third as many output tokens (the launch post says action tokens), and a proof one-seventh the length. No absolute token or cost figure is given. Opus 5.5 in the same setup had completed 26% at 18 hours, when Fable was at 0.86, and barely progressed after. At hour 18 both runs were restarted from their last save with 12 more hours.

The setup

Each agent ran in its own lightweight virtual machine. Anthropic says a single shared machine hit its limits at about 100 agents. Every agent had:

  • a shell and a file editor, with no network access;
  • Lean's toolchain and the Mathlib library;
  • its own Git branch, with the ability to open merge requests into main, where a grader checked each proof;
  • a shared Postgres database, reached from the shell.

The only structure was one lead agent and 999 helpers. The lead was told to split the problem and assign parts; helpers were told to expect work from the lead and ask when unclear. The harness also sent the lead 19 automatic reminders of the time remaining.

How the agents organized themselves

In its first 10 minutes the lead created five database tables: announcements, assignments, an open notes board, a subtasks list any agent could add to, and a log of finished work. It assigned crews of 16 or 17 helpers to each statement and appointed captains. It then asked everyone to judge their statement's difficulty. About 27 minutes in, three statements were already proved. The lead set aside the 47 statements that most helpers had judged infeasible within budget and piled the team onto the remaining 10, growing crews to between 46 and 139 helpers. Those were all proved in about 70 minutes, and the lead moved helpers back to the statements it had shelved.

Around hour 18, what remained was Langlands-Tunnell, a classical result Wiles's proof depends on, the statements resting on it, and one other statement that the team proved along the way. Earlier, the Langlands-Tunnell captain had estimated the proof would take at least 0.7 to 1 million lines of Lean, with less than a 1% chance of finishing in time. The lead moved about 750 helpers onto it, and the team built a tree of 2,418 tasks up to 10 levels deep. Langlands-Tunnell was proved at hour 28.4 and the final theorem at hour 28.7. The final tally counts 54 of 60 statements, because five alternatives were no longer needed and one statement was dropped.

Safety controls

Anthropic describes three layers. Sandboxes had no network, a locked-down kernel and no route to root, with sudo removed. Agents coordinated only through Git and the database. A prompted classifier checked every shell command and file edit before it ran, and an offline monitor reviewed actions afterwards; a block on a helper would have removed just that helper, a block on the lead would have ended the run. Across more than a million tool calls, none met the bar to be blocked. This run used only an action-based monitor. Anthropic has since added one that reads the agents' reasoning; replayed against it, only one call, a check for root access, came close to being blocked.

Limits Anthropic notes

The authors say the result does not remove the need for specialist platforms like Prove2Me, which may still help agents compile Lean code faster and reuse human results. They also note that agents misjudged difficulty and tended to give up on hard problems; most of the 47 shelved statements were proved in the end. Human encouragement at those moments could help. It is a single run per model, reported by Anthropic, on a proof that had already been formalized once, and the posts do not link the resulting proof.

Our take

For builders, the reusable part is the coordination design Anthropic chose: a plain SQL database, picked because its transactions handle concurrent writes from 1,000 agents, one lead, and a task tree that grows when a piece is too hard. Whether that carries over to other long, structured jobs is our speculation, not something Anthropic tested. Also notable: the agents' own difficulty estimates were badly wrong, and the harness sent the lead 19 time-remaining reminders to keep it going.

Source: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (Anthropic, November 9, 2026), Safely scaling multi-agent teams with Claude Fable 5.2 (Claude developer blog, November 9, 2026)

This story first appeared in the claudefox daily briefing for November 10, 2026.