Claude Automates Code Verification, Cutting Proof Costs by Millions of Tokens

Anthropic has demonstrated that its Claude model can now autonomously generate formal proofs of software correctness, a capability that could fundamentally alter the cost structure of quality assurance in the tech industry. By leveraging its Frontier Red Team in partnership with Theorem, the company has shown that AI can not only write code but also mathematically verify its safety, moving the industry closer to a future where automated verification is as standard as unit testing.

The scale of this automation is significant. Over the course of a few weeks, Claude produced approximately 180 million lines of mathematical proof code. This output was derived from test programs within 123 open-source libraries sourced from Google’s OSS-Fuzz security project. The process involved translating crash tests into mathematical problems and analyzing compiled machine code to produce extensive proofs. An independent verification program named Rocq automatically checks these proofs, a workflow that Anthropic describes as requiring minimal human intervention, labeling it "fully autonomous" proof generation despite the human-built foundational tools provided by Theorem.

The financial and operational implications of this efficiency are substantial. Historically, formal verification was prohibitively expensive, reserved for critical systems like Airbus avionics. For context, a team of specialists once required 20 person-years to formally verify the small operating system kernel seL4. By contrast, Claude handled the verification for 123 libraries in weeks. While individual proofs can be computationally intensive, running up to hundreds of millions of tokens, the overall approach reduces the labor costs associated with manual review. Furthermore, the efficiency of newer models is improving; Opus 5.5 solves half the proofs with approximately 40 percent fewer tokens than its predecessor, Opus 5.

Key figures in this development include the researchers at Anthropic’s Frontier Red Team and their collaborators at Theorem. Theorem provided the formal foundations and tools necessary for the verification process, while Anthropic’s team directed the AI model. The partnership highlights a shift in how specialized technical tasks are executed, with AI handling the heavy lifting of proof generation while humans focus on high-level architectural decisions. The project also involved the use of Rocq, an independent verification program, to ensure the integrity of the generated proofs without relying solely on the model that created them.

The practical stakes of this technology lie in its ability to uncover bugs that traditional testing methods miss. During the verification process, Claude identified hundreds of bugs and was authorized to fix them. Many of these defects only manifest under highly specific conditions, such as unusually large input files or particular sequences of memory accesses. In one notable instance, the team found as many security bugs in a single week for a widely used PDF reader as are typically reported for that software in an entire year. In another case, the model caught a logical error that had remained undetected for over a decade.

However, the scope of this verification is defined by strict technical constraints. Anthropic notes that "formally verified" does not equate to "error-free" in all contexts. The proofs hold only under specific conditions, including a defined processor architecture, single-threaded execution, and inputs smaller than 64 MB. These limitations mean that while the method is powerful for specific components, it cannot yet account for the full complexity of modern, multi-threaded applications running on diverse hardware configurations.

Despite these limitations, the project suggests a broader trajectory for software automation. If AI can verify its own code, developers can place greater trust in AI-generated software, reducing the time and capital spent on manual review. This efficiency gain extends beyond verification to the entire coding lifecycle. The reusability of proofs, demonstrated in smaller test runs, further supports the economic viability of this approach. As models become more efficient, the cost of assurance continues to decline, making formal verification a more accessible tool for a wider range of software projects.

The long-term vision for this technology is ambitious. Anthropic expects formal verification with AI to soon become as routine as standard testing. Theorem’s roadmap already targets critical infrastructure components, including glibc, OpenSSL, and cURL. If successful, this shift would answer the question of whether AI can prove the safety of the code it writes, a milestone that could extend the impact of artificial intelligence well beyond simple coding assistants into the core of software reliability and market confidence.