New AI Method Cuts Software Verification Costs

For businesses relying on complex software systems, the cost of ensuring code reliability is finally starting to drop. A new collaboration between Anthropic and Theorem has demonstrated that artificial intelligence can now mathematically verify existing software, a task that previously required years of specialized human labor. This shift suggests that companies may soon spend significantly less on quality assurance while catching critical bugs that traditional testing methods often miss.

The core value of this approach lies in its ability to automate a historically expensive process. Formal verification, which proves that software runs without errors under all conditions, was once reserved for high-stakes industries like aviation. For example, a team of specialists took twenty person-years to verify a small operating system kernel. By using Claude to handle the proof generation, Anthropic and Theorem compressed this timeline to just a few weeks, making the technology more accessible for a broader range of commercial applications.

The scale of the recent project highlights the potential for efficiency gains. The AI model produced approximately 180 million lines of proof code for 123 open-source libraries from Google's OSS-Fuzz security project. This volume of work would have been impossible for a human team to complete in the same timeframe. The process involves translating crash tests into mathematical problems and analyzing compiled machine code, allowing the system to operate with minimal human intervention.

Beyond speed, the method has proven effective at uncovering deep-seated defects. During the verification process, Claude identified hundreds of bugs that had survived years of conventional testing. These errors often only appear under specific conditions, such as unusually large input files or particular access sequences. In one notable instance, the team found as many security bugs in a widely used PDF reader in a single week as are typically reported for that software in an entire year.

The ability to fix these issues automatically further reduces labor costs. The researchers allowed Claude to correct the bugs it discovered, including a logical error that had remained undetected for over a decade. Even when a complete proof fails, the attempt surfaces real problems. This suggests that developers can trust AI-generated software more readily, reducing the time and money spent on manual code review across the entire development lifecycle.

However, the technology comes with specific limitations that businesses must consider. Anthropic notes that formal verification does not mean the software is entirely error-free. The proofs are valid only under strict conditions, including a specific processor architecture, single-threaded programs, and inputs under 64 MB. Since most commercial software runs on varied hardware and configurations, these constraints mean the method is not yet a universal solution.

The compute requirements also remain a significant factor in the cost analysis. Individual proofs can consume hundreds of millions of tokens, and Anthropic has not disclosed the total project costs. Despite this, newer models are improving efficiency. A smaller test run showed that Opus 5.5 can solve half the proofs with about 40 percent fewer tokens than its predecessor, Opus 5. This trend indicates that the cost per verification may continue to decline as model efficiency improves.

The project signals a broader shift in how software automation is approached. If AI can write code and guarantee its correctness under defined conditions, the focus moves from basic programming to safety assurance. This could have consequences that extend well beyond current coding tools, potentially transforming how critical infrastructure is managed and maintained.

Looking ahead, the roadmap for this technology targets major components of the software stack. Theorem plans to apply this method to critical infrastructure libraries such as glibc, OpenSSL, and cURL. Anthropic expects that formal verification with AI will soon become as routine as standard testing. For New England businesses and their partners across the region, this evolution suggests a future where software reliability is built in through automated proofs rather than after-the-fact testing.