Vals raised $40M to sell AI benchmarks vendors can't game

Vals closed a $40 million Series A led by Andreessen Horowitz in August, after being seeded by 8VC and Bloomberg Beta, per TechCrunch. Founded in 2024, the company runs AI evaluations in law, finance, coding, mental health, cybersecurity, biosecurity and the law of armed conflict, and it added a federal agency program.

The pitch is simple once you sit with it: a benchmark stops being useful the moment it's published. Publish the questions and answers, and any lab can tune a model to ace that specific test without getting better at the underlying task. Vals' fix is to never publish. Test materials stay private, so a model either handles the real problem or it doesn't.

That's a direct shot at the leaderboards driving most model comparisons. Public leaderboards are readable by definition, so every score on them is a score a vendor has had the chance to optimize against. If Vals is right that private evals are the only trustworthy kind, those open scores measure test-taking skill, not capability.

The growth numbers back up demand for that pitch. Revenue is eight times what it was a year ago. Headcount went from 8 to 25, with 10 to 15 more hires planned. Companies buying private evals, especially in regulated fields like finance and cybersecurity, are the ones with the most to lose from a model that scores well on a public test and fails on their actual workload.

What the TechCrunch piece doesn't cover is how Vals verifies its own evaluators aren't the ones leaking test content, or how a private benchmark stays credible to outsiders who can't audit it. Trusting a black-box scorer is a different problem than trusting a public one, just a quieter version of it.