a16z Enters AI "Third-Party Oracle" Space: Vals Raises $40 Million

According to Dynamic Beating monitoring, AI model evaluation company Vals AI has completed a $40 million Series A funding round, valuing the company at $400 million, led by a16z. Existing shareholders such as 8VC, Pear VC, and Bloomberg Beta also participated.

Vals focuses on third-party evaluation of large models. Currently, many models are released with performance metrics tested and reported by the manufacturers themselves. In contrast, Vals conducts unified testing of different models using real-world tasks in programming, finance, law, healthcare, and more. The company claims that its results have been referenced by OpenAI, Anthropic, Google, Meta, and xAI models, and its revenue for this year has already reached eight times that of the full year 2025.

a16z is interested in the value of this "third-party adjudication" layer. Public benchmarks are becoming increasingly vulnerable to score manipulation, data leakage, and even targeted optimizations by manufacturers, where a high leaderboard score does not necessarily equate to actual usability.

Additionally, Vals has now fully launched Vals Smith, allowing any GitHub repository to be turned into a Coding Benchmark. The system extracts real development tasks from historical pull requests and then uses hidden tests to check whether the model can complete them. This enables companies to directly test which model is best suited for their codebase, rather than relying solely on publicly available leaderboards.