Software engineering has long accepted a fundamental bottleneck: verifying code correctness demands astronomical labor costs from scarce specialists. Corporate product budgets burn for years on manual testing that routinely misses critical logic flaws. Anthropic is attempting to break through this financial and temporal barrier, partnering with Theorem to deploy the Claude model for mathematical verification of existing software.
Anthropic's Frontier Red Team, working with partner Theorem, let Claude take over the proof work instead.
Such performance alters the economics of quality control, shifting routine verification from elite consultancy services to automated pipelines. The independent Rocq verifier handles the automated checking of these proofs, and Anthropic notes that the process requires virtually no human verifier involvement.
Scaling this approach runs straight into rigid technological walls that directly impact operational expenditure and token economics. The generated correctness proofs hold only under strict operational boundaries: fixed processor architectures, single-threaded execution, and input limits capped at 64 megabytes. Real-world enterprise systems running across diverse hardware and multi-threaded environments demand entirely different orders of verification, exponentially increasing token expenditure as codebases scale.
Compute economics in this domain dictate that long-term enterprise software maintenance budgets will face severe structural shifts. As engineering teams attempt to apply mathematical verification to sprawling legacy systems, the token cost per verification cycle multiplies rapidly due to state-space explosion. However, engineering optimization shows measurable progress: newer model iterations handle the workload more efficiently, with Opus 5.5 resolving half of the proofs while consuming roughly 40 percent fewer tokens than Opus 5. This reduction in token expenditure directly targets the primary barrier keeping formal verification out of standard enterprise software maintenance budgets.
The practical utility of the approach demonstrates itself through the discovery of latent defects in industrial software components. During the verification of a widely deployed PDF reader, the team uncovered as many security vulnerabilities in a single week as typically surface in the codebase over an entire year. In another instance, the model exposed a logical flaw that had remained hidden for more than a decade.
Anthropic expects formal verification with AI will soon become as practical as standard testing, and Theorem's roadmap already targets critical infrastructure like glibc, OpenSSL, and cURL.
Such outcomes lay the groundwork for moving formal verification into the category of standard testing routines, fundamentally altering long-term enterprise software maintenance budgets by catching critical flaws before production deployment. When development teams acquire instruments capable of guaranteeing code safety under specified parameters, the entire paradigm of risk assessment for generative systems inside enterprise products shifts from reactive patching to proactive mathematical certainty.
High token costs and harsh architectural constraints still leave this methodology inside the zone of expensive experiments rather than everyday commercial deployment, forcing decision-makers to carefully balance compute economics against the high cost of undetected software vulnerabilities.
Editorial analysis
How the story developed
Anthropic first slashed prices for its lightweight Claude Haiku 5.5 model to $0.10 per million input tokens and $0.50 per million output tokens on 18 October, and shortly after reported a 90 to 99 percent cut in data harvested via distillation attacks on 22 October. Now, the company tackles software engineering's manual bottleneck by partnering with Theorem to deploy Claude for formal mathematical verification of software. Anthropic continues tackling core developer bottlenecks by embedding Claude directly into critical software verification workflows.
Sources: our stories of 18.10.2026, 22.10.2026