Claude Proves Code Safety, Cutting Verification Costs

For decades, proving that software is perfect has been a luxury reserved for the most critical systems. It was too expensive and too slow for the average developer. Now, that barrier has collapsed. Anthropic has demonstrated a way to use its Claude model to mathematically prove that existing code is correct, automating a process that once required armies of specialists.

The scale of this new capability is striking. In just a few weeks, Claude generated roughly 180 million lines of mathematical proof code. This work covered 123 open-source libraries from Google’s OSS-Fuzz security project. The effort was a collaboration between Anthropic’s Frontier Red Team and its partner, Theorem. Instead of writing new software from scratch, the AI focused on verifying code that already existed.

The method involves translating crash tests into mathematical problems. Claude then analyzes the compiled machine code to produce extensive proofs. An independent program called Rocq automatically checks these proofs for validity. Anthropic describes the process as nearly autonomous, requiring very few human reviewers. While the underlying tools and formal foundations were built by humans at Theorem, the heavy lifting of proof generation was handled by the model.

The immediate benefit is the discovery of flaws that traditional testing misses. During the verification process, Claude identified hundreds of bugs. Many of these errors only appear under specific, rare conditions, such as unusually large input files or specific sequences of data access. Because they are hard to trigger, they often slip past standard quality control measures.

One example highlights the speed of detection. While verifying 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 case, Claude caught a logical error that had gone undetected for more than a decade. Even when a complete proof fails, the attempt still surfaces real issues, according to Anthropic.

However, the term “formally verified” carries strict conditions. It does not mean the software is error-free in every possible scenario. The proofs only hold under specific constraints, such as a particular processor architecture, single-threaded programs, and input files smaller than 64 MB. In real-world use, these libraries often run on different hardware and configurations, so the verification is not a universal guarantee.

The computational cost remains high. Individual proofs can require hundreds of millions of tokens to complete. Anthropic has not disclosed the total cost of the project. Despite this, the efficiency gains are significant. Newer models are solving problems with less data; Opus 5.5, for instance, solves half the proofs with about 40 percent fewer tokens than its predecessor.

This shift changes the question for software development. It is no longer just whether AI can write code, but whether it can prove that code is safe. If the answer is yes, the implications extend beyond simple coding tools. Anthropic expects formal verification to become as routine as standard testing. Theorem’s roadmap already targets critical infrastructure libraries like glibc, OpenSSL, and cURL, signaling a push toward securing the backbone of the internet.

For New England businesses relying on digital infrastructure, this technology could mean more reliable software with less human oversight. The ability to automate safety proofs reduces the time and cost associated with reviewing AI-generated code. It suggests a future where software reliability is built in, not just tested after the fact.