180 Million Lines: Claude's Massive Code Verification Push

Software developers have long faced a difficult choice. They can write code quickly using standard tools, but they cannot easily prove that it is perfect. Proving that software runs without errors under every possible condition has always been incredibly expensive. In the past, this level of rigor was reserved for high-stakes systems like airplane avionics. Now, that balance is shifting. Anthropic claims its AI model, Claude, can now write code and prove its safety automatically, reducing a task that took years of human effort to a matter of weeks.

The challenge of formal verification is well known. It is a mathematical process that checks if code meets specific requirements. It is slow and costly. For example, a team of specialists spent 20 person-years to verify the seL4 operating system kernel. Because of this high cost, most everyday software has never been formally verified. It relies on testing, which only checks for known problems and misses rare edge cases. This leaves a gap in security that attackers often exploit.

To address this, Anthropic’s Frontier Red Team partnered with Theorem. They used Claude to perform formal verification on existing software. The model did not write new applications. Instead, it analyzed compiled machine code from 123 open-source libraries in Google’s OSS-Fuzz security project. Claude translated crash tests into mathematical problems and then generated proofs that the code behaves correctly. This work produced roughly 180 million lines of proof code in just a few weeks.

The process is largely autonomous. An independent program called Rocq checks the proofs generated by Claude. According to Anthropic, this requires very little human review. The researchers describe the system as fully autonomous, although the underlying tools and formal foundations were built by humans at Theorem. The goal is to make formal verification as routine as standard testing, allowing developers to trust AI-generated software more confidently.

The results have already shown significant value. While generating proofs, Claude identified hundreds of bugs that conventional testing had missed. Many of these errors only appear under specific conditions, such as unusually large input files or particular access sequences. In one instance, the team verified a widely used PDF reader. They found as many security bugs in one week as are typically reported for that software in an entire year.

In another case, Claude caught a logical error that had remained undetected for over a decade. The model was even allowed to fix some of these bugs itself. Anthropic notes that even when a complete proof fails, the attempt often surfaces real issues. This means the process adds value regardless of the final outcome. It helps developers find problems before they reach users.

However, there are strict limits to what this technology can do. Anthropic clarifies that formally verified does not mean error-free. The proofs only hold under specific conditions. These include a specific processor architecture, single-threaded programs, and inputs under 64 MB. In real-world use, these libraries run on various hardware and configurations. Therefore, the verification applies to a defined scope rather than every possible scenario.

The cost of computing is also a factor. Individual proofs can consume hundreds of millions of tokens. Anthropic has not reported the total project costs. However, the approach offers potential savings in the long run. If developers can trust AI-verified code, they spend less time reviewing it. This reduces costs across the entire coding process. Smaller tests also show that proofs are reusable and that newer models are more efficient. Opus 5.5 solves half the proofs with about 40 percent fewer tokens than Opus 5.

This project changes the conversation around AI in software. The question is no longer if AI can program. It is whether AI can prove that its code is safe. If the answer is yes, the implications could extend far beyond current coding tools. Anthropic expects formal verification with AI to become as practical as standard testing soon. Theorem’s roadmap already targets critical infrastructure components like glibc, OpenSSL, and cURL. This suggests a future where mathematical proof is a standard part of software development.

For now, the technology remains in the hands of specialized teams and partners. But the direction is clear. As models become more efficient and costs drop, formal verification may move from the edge of software engineering to the center. Developers in New York and New Jersey, like others, may soon rely on AI not just to write their code, but to guarantee its safety. The shift from testing to proving marks a significant step in the automation of software.