Andreessen Horowitz Backs Vals With $40M to Reset AI Benchmarks
Vals raises a $40 million Andreessen Horowitz-led Series A to build private, industry-specific AI benchmarks for companies and federal agencies at scale.
Summary
Founded in 2024, Vals is building private AI evaluations after 25-year-old co-founder Rayan Krishnan, a former Palantir intern who worked for Microsoft and Stanford’s AI lab, concluded that academic benchmarks lagged fast-improving models. Because favorable scores support marketing, public tests invite developers to train models against known questions. Vals withholds its materials and assesses human-quality work in law, finance and coding, plus beneficial and harmful outcomes involving recursive self-improvement, mental health, cybersecurity, biosecurity and applying the Geneva Convention under the law of armed conflict.
Vals secured a seed round led by 8VC and Bloomberg Beta in 2025, followed by a $40 million Series A led by Andreessen Horowitz in August 2026. By September 19, 2026, revenue had reached eight times its year-earlier level, while headcount had risen from eight at the start of the year to 25. Vals plans a larger office, 10 to 15 additional hires and has launched an evaluation program for federal agencies.
Model developers pay Vals to identify weaknesses and improve systems, a model Krishnan compares with students paying the College Board for the SAT; prospective buyers increasingly use evaluations when selecting AI models. Krishnan expects benchmarks to shape public trust, purchasing, investment plans and corporate filings. He said SpaceX had gone public, Anthropic was slated for later in 2026 and OpenAI could follow soon.
Positives
- $40 million from an Andreessen Horowitz-led Series A gives Vals capital to expand its private AI evaluation platform.
- Eightfold year-over-year revenue growth indicates rising demand for model testing and procurement guidance.
- Headcount increased from eight to 25 in 2026, with another 10 to 15 hires planned.
- Private test materials reduce opportunities for developers to train models against known benchmark questions.
- Federal agencies can now access Vals evaluations through the startup’s newly launched government program.
Risks & concerns
- Legacy academic benchmarks are failing to keep pace with rapidly improving AI models.
- Publicly available tests allow developers to optimize models for known questions, weakening scores as evidence of real capability.
- Unchecked models could create harms in mental health, cybersecurity, biosecurity and armed conflict, areas Vals is expanding its tests to examine.