AI Breakthrough: Claude Generates 180 Million Lines of Code Proof in Weeks

Anthropic’s Claude AI model has successfully generated approximately 180 million lines of mathematical proof code in a matter of weeks, marking a significant leap in the ability to automatically verify the safety of existing software. This achievement, accomplished in partnership with Theorem, shifts the focus of artificial intelligence from merely writing code to mathematically proving that the code functions correctly under defined conditions.

The process, described by researchers as fully autonomous, involved Claude analyzing compiled machine code from 123 open-source libraries within Google’s OSS-Fuzz security project. Instead of creating new software, the model translated existing crash tests into mathematical problems. An independent verification program named Rocq then automatically checked these proofs, a workflow that Anthropic states requires minimal human intervention despite the formal foundations being built by human experts at Theorem.

Historically, formal verification has been reserved for high-stakes environments like Airbus avionics due to its prohibitive cost. For context, a team of specialists previously required 20 person-years to formally verify the small operating system kernel seL4. By offloading this labor-intensive work to Claude, Anthropic has demonstrated that the barrier to entry for rigorous software safety checks can be lowered dramatically, potentially making such verification as routine as standard testing.

A critical benefit of this automated approach is its capacity to detect flaws that conventional testing methods often miss. During the proof generation process, Claude identified hundreds of bugs and was permitted to fix them independently. These errors were often elusive, manifesting only under specific conditions such as unusually large input files or particular sequences of data access, which typically allow them to slip past standard quality assurance protocols.

The efficiency of the system was starkly illustrated during the verification of a widely used PDF reader. In a single week of analysis, the team discovered as many security bugs as are typically reported for that specific software over an entire year. In another instance, the model caught a logical error that had remained undetected for more than a decade, proving that the attempt at formal verification surfaces real-world issues even when a complete proof is not immediately achieved.

Despite these successes, Anthropic maintains that formal verification does not equate to absolute error-free status. The proofs generated by Claude are valid only under strict conditions, including specific processor architectures, single-threaded program execution, and input sizes under 64 MB. Since these libraries often run on diverse hardware and configurations in practical applications, the verification covers a defined subset of operational parameters rather than every possible scenario.

The computational demands of the project are considerable, with individual proofs consuming hundreds of millions of tokens. While Anthropic has not disclosed the total project costs, the company notes that newer models are becoming more efficient. Specifically, the Opus 5.5 model solves half the proofs using approximately 40 percent fewer tokens than its predecessor, Opus 5, suggesting that the economic viability of this technology is improving rapidly.

The broader implication of this work is a shift in the question from whether AI can program to whether it can prove the safety of its own output. If developers can trust AI-generated software more thoroughly, they can spend less time on manual review, saving costs across the entire coding lifecycle. Anthropic expects this method to become standard practice soon, with Theorem’s roadmap already targeting critical infrastructure components like glibc, OpenSSL, and cURL for verification.