Anthropic shows 1,000 Claude Fable 5.2 agents reproduce Fermat's Last Theorem proof in 28.7 hours

Anthropic says a team of 1,000 Claude Fable 5.2 agents reproduced a Lean proof of Fermat's Last Theorem in 28.7 hours, using far fewer resources than a prior human-guided attempt, with zero actions blocked across over 1 million tool calls.

Anthropic published a post on its developer blog on November 9, 2026, describing an experiment in which a team of 1,000 Claude Fable 5.2 agents, one lead agent plus 999 helper agents, autonomously reproduced a Lean formalization of Fermat's Last Theorem. According to the post, titled Safely scaling multi-agent teams with Claude Fable 5.2, the agents finished the task using layered safety controls and far fewer resources than a prior human-guided attempt.

The numbers behind the run

One lead agent coordinated 999 helper agents, for a total of 1,000 agents, to reproduce the proof
The task was completed in 28.7 hours
The run used about one third of the output tokens of a prior September effort, and the resulting proof was about one seventh as long
The September effort took 11 days, about 6 billion output tokens, and 13 million lines of Lean code, with direct human guidance
Helper agents judged 47 of the initial 60 theorem statements infeasible within budget, and the lead agent created 5 coordination database tables within the first 10 minutes without specialized scaffolding
Safety layers included sandboxed virtual machines with no network access, a real time command blocking monitor, and offline review

What it means for people who build with AI

Anthropic frames the run as a concrete, measured data point on orchestrating agent swarms at a scale of 1,000 agents together with the safety tooling needed to do so safely, something directly relevant to developers building multi-agent coding and research systems. The post notes that the hardest sub-problem the agents tackled, known as Langlands-Tunnell, generated 2,418 tasks spread across 10 levels of depth, showing how deep the coordination tree grew during the run. Anthropic also reports that across more than 1 million tool calls made during the run, zero were blocked by the safety monitor.

Sources

Safely scaling multi-agent teams with Claude Fable 5.2 (Anthropic (claude.dev developer blog), November 9, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.