AI benchmarking startup Vals is betting that private, task-specific evaluations can become infrastructure for an industry that has outgrown many of its standard tests. The company says it wants to measure what models can accomplish in real work, as well as what can go wrong when increasingly capable systems operate with less supervision.
The pitch has attracted substantial funding. Formed in 2024, Vals raised a seed round led by 8VC and Bloomberg Beta, followed by a $40 million Series A led by Andreessen Horowitz. Its growth claims and long-term ambitions come from the company and have not been independently verified in the TechCrunch report.
Private Tests Aim to Reduce Benchmark Gaming
Co-founder Rayan Krishnan argues that academic benchmarks have not kept pace with rapidly improving models. Many established tests are public, allowing developers to train against their questions or otherwise optimize for a familiar scoring system. A high score can then become a marketing asset without necessarily proving that a model performs reliably in practice.
Vals keeps its specific test materials private. Instead of concentrating on abstract knowledge or exam-style questions, it evaluates complex tasks in fields such as law, finance and coding. Krishnan describes the goal as determining whether a model can produce work comparable in quality to a human’s output within a domain.
That design addresses one weakness in public benchmarks, but secrecy creates its own trade-off. Outside researchers cannot fully inspect private tasks, and customers must trust the evaluator’s methods and controls. TechCrunch does not provide independent performance comparisons between Vals and other evaluation systems.
Evaluations Extend From Performance to Risk
Vals is expanding beyond conventional workplace tasks. Krishnan said the company has developed or is developing benchmarks involving recursive self-improvement, mental health, cybersecurity, biosecurity and the law of armed conflict. One effort examines whether models can apply the Geneva Conventions.
The company says evaluations should capture negative outcomes as well as useful capabilities. That means asking not only whether a model completes a task, but also what damage could follow if it acted incorrectly or without adequate controls.
AI companies pay Vals to test their own models. Krishnan compares the arrangement with students paying an examination body, arguing that credible measurement helps developers identify weaknesses and helps buyers choose among systems. The commercial relationship also requires clear safeguards against conflicts of interest: a customer may want useful criticism while also caring about how results affect its reputation. The source does not describe Vals’s publication rules or governance for disputed findings.
Funding Supports a Bid for Industry-Wide Influence
Vals says its revenue is eight times higher than last year. Its staff grew from eight people at the beginning of the year to 25, and Krishnan plans to hire another 10 to 15 employees and move to a larger office. The company has also launched a program to provide model evaluations to federal agencies.
Krishnan expects benchmarks to become important in procurement, investment decisions and eventually public-company disclosures as AI becomes more deeply embedded in the economy. That is a projection rather than an established reporting requirement.
The opportunity rests on a real credibility problem: model makers benefit when tests produce strong scores, while widely known benchmarks can lose value as training targets. Vals’s answer is to professionalize evaluation through controlled test sets and domain expertise.
Its challenge will be proving that private measurement deserves public trust. Funding, customer growth and broader subject coverage can establish scale, but a benchmark becomes authoritative only if buyers, researchers and affected institutions accept its independence and methodology. Vals is positioning itself for that role while the standards for judging evaluators remain unsettled.
