Anthropic's Claude: 180 Million Lines of Proof Code

Anthropic's Claude has generated approximately 180 million lines of mathematical proof code across 123 open-source libraries within a period of only a few weeks. The company announced this development today in conjunction with its partner, Theorem, detailing a joint effort to mathematically verify existing software code. This volume of output highlights a significant shift in the approach to software verification, a process Anthropic describes as having been historically prohibitively expensive. The company's Frontier Red Team led the project, utilizing Claude to perform proof generation on existing codebases.

The core of this initiative is the application of Claude to the task of formal verification, a method used to prove software runs error-free under all circumstances. Anthropic and Theorem worked together to automate this process. The model translated crash tests from Google's OSS-Fuzz security project into mathematical problems, analyzed the compiled machine code, and produced extensive proofs of its correctness. An independent verification program called Rocq automatically checks these proofs, and Anthropic states that the process requires almost no human oversight.

This effort targets the high cost associated with traditional formal verification methods. For instance, a team of specialists required 20 person-years to formally verify the seL4 operating system kernel. This expense has historically limited the use of formal verification to extreme cases, such as Airbus avionics software. By automating the generation of proofs with Claude, Anthropic aims to make formal verification as routine as standard testing.

During its proof-generation work, Claude identified hundreds of bugs that were missed by conventional testing. The researchers state these bugs only manifest under very specific conditions, such as with unusually large input files or a particular sequence of accesses. In one case involving a widely used PDF reader, the team found as many security bugs in a single week as are typically reported for that software in an entire year. In another instance, Claude discovered a logical error that had gone undetected for over a decade.

However, Anthropic acknowledges that “formally verified” does not mean “error-free.” The proofs are only valid under strict conditions: a specific processor architecture, single-threaded programs, and inputs under 64 MB. In practice, these libraries run on different hardware and in different configurations. The compute cost for this work is also substantial, with individual proofs running up hundreds of millions of tokens. Anthropic did not report total project costs for the effort.

The project's success has implications for the software automation market. If AI can write code and guarantee its correctness under defined conditions, the question shifts from whether AI can program to whether it can prove its code is safe. This capability could reduce the time developers spend reviewing AI-generated software, lowering costs across the coding process. Anthropic also notes that proof generation is becoming more efficient, with Opus 5.5 solving half the proofs with about 40 percent fewer tokens than Opus 5.

Anthropic expects formal verification with AI will soon become as practical as standard testing. The company's partner, Theorem, has a roadmap that targets critical infrastructure software including glibc, OpenSSL, and cURL. The potential to write and automatically verify code could therefore extend beyond current coding tools to impact a much broader range of software development and security.