180 Million Lines: Anthropic's Claude Automates Formal Software Verification

The cost of ensuring software integrity has historically been a significant barrier to widespread adoption of rigorous safety standards. On October 23, 2026, Anthropic announced a breakthrough in this area, revealing that its Claude model can now mathematically verify the correctness of existing software code with minimal human intervention.

The initiative, conducted in partnership with Theorem, leverages Claude to generate formal proofs that software behaves as intended. In a matter of weeks, the AI produced approximately 180 million lines of proof code for 123 open-source libraries. This volume of work, which would have taken human specialists years to complete, highlights a potential shift in how the software industry approaches quality assurance and risk management.

Traditionally, formal verification has been reserved for high-stakes environments due to its prohibitive labor costs. For example, verifying the seL4 operating system kernel required 20 person-years of specialist effort. By automating this process, Anthropic aims to make such rigorous checks as routine as standard unit testing, thereby reducing the overall cost of software development and maintenance.

The methodology does not involve writing new code from scratch. Instead, Claude translates existing crash tests from Google’s OSS-Fuzz security project into mathematical problems. The model then analyzes compiled machine code to generate extensive proofs, which are subsequently checked by an independent verification program called Rocq. Anthropic describes this workflow as “fully autonomous,” noting that it requires very few human reviewers, although the underlying formal foundations were established by human experts at Theorem.

Beyond generating proofs, the process proved effective at identifying latent defects. During the verification of the 123 libraries, Claude detected hundreds of bugs that conventional testing methods had missed. These errors often manifested only under specific conditions, such as unusually large input files or particular access sequences. In one notable instance, the team found as many security vulnerabilities in a single week within a widely used PDF reader as are typically reported in an entire year. In another case, a logical error that had remained undetected for over a decade was uncovered.

However, the approach comes with strict practical limitations. Anthropic clarifies that “formally verified” does not equate to “error-free” in all real-world scenarios. The proofs are valid only under specific conditions, including a defined processor architecture, single-threaded execution, and input sizes under 64 MB. Since these libraries often run on diverse hardware configurations, the verification covers a specific subset of operational environments.

The economic implications of this technology are significant, particularly regarding compute costs. Individual proofs can consume hundreds of millions of tokens, though Anthropic has not disclosed the total project expenditure. Despite this, the efficiency gains are evident in model performance; newer versions like Opus 5.5 require roughly 40 percent fewer tokens to solve half the proofs compared to Opus 5. This trend suggests that as AI models become more efficient, the cost of rigorous verification will continue to decline.

The long-term outlook for the industry is shaped by the potential for reusing these proofs. A smaller test run indicated that generated proofs are reusable, which could further reduce costs for developers who need to verify code across different contexts. If AI can both write and prove the safety of code, the focus of software development may shift from basic functionality to guaranteed correctness under defined parameters.

Anthropic and Theorem see this as a major step toward broader software automation. The partner’s roadmap already targets critical infrastructure components such as glibc, OpenSSL, and cURL. As these tools become more practical and cost-effective, the question for the market is no longer whether AI can program, but whether it can reliably guarantee that its output is safe, a capability that could transform the entire software lifecycle.