Thousand Anthropic AI Agents Formalize Proof of Fermat's Last Theorem in 30 Hours
The company announced that a thousand Claude Fable 5.2 agents recorded a proof of Fermat's Last Theorem in Lean within 30 hours. The experiment also involved Claude Opus 5.5.
Anthropic has updated the Claude Fable 5.2 and Claude Mythos 5.2 models. In an experiment with Fable 5.2, a thousand agents received initial instructions and a list of intermediate statements, then independently distributed the work on the proof in Lean. According to the company, human intervention was rare, and the agents were not provided with mathematical reference materials.
For comparison, Anthropic used approximately a thousand Claude Opus 5.5 agents with the same configuration. In the first 18 hours, the model completed 26% of the formalization and made almost no progress in the next 12 hours. Both teams continued working from a saved state after 18 hours. The graph shows one run of each model.
On an early version of Fable 5.2, a team of agents also migrated a critical service from Python to Rust. According to Anthropic, its startup time was reduced by approximately 26 times. In another experiment, over a thousand sub-agents created a prototype operating system from scratch in eight days, which boots on a laptop.
On FrontierCode 2.0, Fable 5.2 scored 49.1% at the Xhigh effort level, compared to 32.4% for Fable 5.1. In High mode, it demonstrated roughly the same performance as Opus 5.5 in Max mode, at approximately half the cost per task.
Fable 5.2 and Mythos 5.2 use the same base model. Fable is available to all users and has additional restrictions for high-risk tasks in cybersecurity and biology. Verified organizations can apply for access to Mythos through Anthropic’s programs. On the Claude Platform API, Fable 5.2 is available under the name claude-fable-5-2. The rate is $8 per million input tokens and $40 per million output tokens, and in Fast Mode, $16 and $80, respectively.
