Can LLMs Finally Scale Formal Verification? Anthropic Bets on 'Proof-at-Scale' — and Developers Aren't Fully Convinced

Anthropic Reframes Formal Verification as an LLM-Native Problem

Anthropic has published research arguing that large language models are shifting the economics of formal software verification, long considered a prohibitively expensive and labor-intensive discipline reserved for safety-critical niches like avionics and cryptography. The core claim is that LLM agents can now generate and iterate on formal proofs at a scale that human proof engineers could never match, potentially bringing mathematical guarantees to mainstream codebases. This is happening now precisely because frontier models have reached a level of code reasoning where they can interact with proof assistants and verification toolchains in a semi-autonomous loop, rather than merely autocompleting snippets. The headline-grabbing piece is a collaboration with Theorem.dev aiming to formally verify the entirety of nixpkgs — the massive Nix package collection — by 2027, an audacious benchmark for the approach. The underlying thesis is that verification, once a bottleneck of human expertise, could become a compute-bound problem that scales with model capability.