Vals

Vals is an independent AI evaluation company founded in 2024 and based in San Francisco. It builds and runs its own benchmarks rather than reporting scores supplied by model vendors, covering economically valuable work such as finance, software engineering and law alongside frontier risks including cybersecurity, recursive self-improvement and mental health. Its founder, Rayan Krishnan, worked at Palantir, Microsoft and Stanford's AI lab before starting the company.

The commercial argument is that a benchmark a laboratory can read is not a measurement. Public test sets leak into training data, whether deliberately or through ordinary web scraping, so a model can score well on a benchmark it has effectively already seen. Vals keeps its test material private and pairs public tasks with held-out equivalents — its CUA-bench agent benchmark, for instance, pairs three named commercial video games with unnamed games of the same genre, so a model trained on the public titles reveals it as a gap between the two halves. Results are published as public leaderboards and written reports that anyone can read, while the evaluation work itself is sold to enterprises and government.

Andreessen Horowitz led a $40 million Series A in August 2026, after a seed round from 8VC and Bloomberg Beta. The company has said revenue is roughly eight times what it was a year earlier and that headcount grew from 8 at the start of 2026 to about 25 by mid-year, with a federal evaluation programme among its more recent work.

Vals sits in a category that matters more as model claims get harder to check independently. Its closest peer in the catalogue is Arena, which crowdsources human preference votes into a public leaderboard, and Patronus AI, which sells evaluation and agent-testing infrastructure to teams building on models. Vals differs from both in running the evaluations itself and withholding the tests.