The round
Vals has raised a $40m Series A led by Andreessen Horowitz, TechCrunch reported on Friday. The company’s earlier seed round was led by 8VC with Bloomberg Beta participating.
Founded in 2024, Vals has grown from 8 people at the start of the year to 25, with plans to add another 10 to 15. Co-founder Rayan Krishnan is 25. Revenue is eight times what it was a year ago.
What it sells
The product is a benchmark that is not public. Vals evaluates models on complex, real-world professional tasks — in law, finance and coding among other domains — and keeps the test material confidential, so a lab cannot fine-tune against the questions.
The measure is whether a model can “do work that produces a product of the same quality as a human”, rather than whether it can answer exam questions. Companies pay to have their models tested, in the way students pay to sit a standardised test — the examiner is funded by the examinee, which is the same structural tension that runs through credit ratings and, for that matter, through this week’s argument over who should evaluate frontier labs.
The domains have expanded well past professional services, and now include recursive self-improvement, mental health, cybersecurity, biosecurity and the law of armed conflict.
Who the customers are
TechCrunch describes the customers as the companies whose models are tested. The published domain list suggests two distinct reasons to buy. A lab wants a score it can quote that its own training run cannot have absorbed. A buyer deploying a model wants to know how it behaves on the work it will actually be given — which is the likeliest explanation for why the list has grown to include biosecurity and the law of armed conflict alongside contract review.
Why private tests are being funded now
Public benchmarks have a contamination problem that has become impossible to ignore. Once a test is on the internet, it is in the next training run, and a score on it stops measuring capability and starts measuring exposure. Held-out, privately administered tests are the obvious correction, and the reason a venture firm will now write a cheque for one.
The cost is verifiability. A benchmark nobody can inspect cannot be audited by anyone outside the company that runs it. Whether a Vals score means what it claims rests on trusting Vals, which is exactly the property public benchmarks were supposed to avoid — and exactly the trade the market appears willing to make.
That trade is being made elsewhere too. MLPerf added agent tests this week, and Anthropic put a consultancy inside its own building to evaluate its models. The common thread is that measurement is becoming a business rather than a public good.
