Vals raises $40 million to fix AI benchmarking, revenue up 8x

Vals was formed in 2024 and built its reputation on a simple complaint: benchmarks that are publicly available can be trained against, so they stop proving much. Co-founder Rayan Krishnan says the academic benchmarks were not keeping up with the frontier of new, capable models, and that benchmarks should verify what companies actually advertise. Vals' answer is to keep its test materials undisclosed and to judge models on work tied to specific industries such as law, finance and coding.

The paying-customer model can look counterintuitive, since a company is buying a test it might fail. Krishnan compares it to a student paying the College Board to sit the SAT, and the article frames the value as troubleshooting and improvement over time, with these evaluations increasingly used by companies deciding which AI models to acquire. To date the company has raised a seed round led by 8VC and Bloomberg Beta, then $40 million in a series A led by Andreessen Horowitz; it also recently launched a program providing model evaluations to federal agencies.

The internal signals are mixed in a way worth watching. Vals says revenue is now eight times last year's level, and headcount has tripled from eight to 25 with plans to add 10 to 15 more. Krishnan argues that as AI companies go public and models become a core part of the economy, evaluations of this kind will shape usage and public filings. Whether that holds may depend on how much weight corporate and government buyers place on an undisclosed test whose methods they cannot inspect themselves.