Anthropic AI Updates: September 5, 2026

1. Claude Agents Formalized Fermat’s Last Theorem in Lean Over 11 Days

Anthropic. A multi-agent harness built on Claude Code produced a machine-checkable Lean formalization of Wiles’s proof of Fermat’s Last Theorem, running autonomously for 11 days. The output is roughly 13 million lines of Lean, more than five times the size of Mathlib, the community’s principal proof library, and contains 30,300 proved theorems of which 29,500 feed the final proof. The run consumed about 6 billion output tokens from an internal research model Anthropic describes as comparable to Claude Fable 5.1. Coordination ran through Prove2Me, a platform designed by Tianyi Peng and collaborators at Columbia University. In a smaller companion experiment, three Claude Max personal subscriptions formalized Vinogradov’s Three Primes Theorem in three days. Source

2. Claude Code v2.1.261 Adds /skill-doctor and Raises Inline Output Limits

Anthropic. The release adds /skill-doctor, which reports which loaded skills go unused and what each costs in context so they can be pruned. Two new settings, bashOutputMaxChars and taskOutputMaxChars, raise how much command and background-task output Claude receives inline before it spills to a file, up to 128K characters. The release also adds --append-subagent-system-prompt-file for subagent system prompts too large to pass on the command line, and an “Organization policy” line in /status and claude doctor explaining why a managed policy failed to load, such as a proxy not passing the endpoint through. On the VS Code side it adds an Add server form and Remove action to the MCP servers dialog, a custom output style walkthrough, session archiving from the right-click menu, and a fold button on permission prompts. Source