Anthropic uses Claude to generate formal proofs for open-source code

Anthropic demonstrated a new application for its Claude model family, directing the system to automatically generate mathematical proofs confirming the functional correctness of existing software. Across tests spanning 123 open-source libraries, Claude produced approximately 180 million lines of formal proof code. The initiative was designed to evaluate whether language models can bridge the gap between traditional unit testing and rigorous mathematical verification, which has historically been constrained by the extensive human labor required to manually draft proof specifications.

During the automated verification process, Claude flagged underlying software defects in multiple packages that had evaded standard testing suites for years. Anthropic stated that its broader development goal is to establish automated formal verification as a routine operational standard alongside conventional software testing. By shifting proof generation to automated agents, the research team aims to eliminate subtle implementation flaws in critical codebases without requiring software engineers to write hundreds of thousands of lines of formal mathematical logic by hand.