Vals disclosed on Sept. 19 that its revenue has grown eightfold over the past year as the AI evaluation company expands private, task-based benchmarks for enterprises and federal agencies. The growth turns a technical problem—how to compare fast-changing models—into a procurement and governance question for companies putting AI into production.
The San Francisco startup told TechCrunch that it has expanded from eight employees at the start of 2026 to 25 and expects to add 10 to 15 more. The profile follows a $40 million Series A led by Andreessen Horowitz in August. Vals also said it has launched a program that provides model evaluations to federal agencies.
What Vals says it is measuring
Public AI leaderboards often rely on academic questions or openly available datasets. That creates two related problems: models can absorb test material during training, and developers can optimize systems for a known exam without proving that the same systems perform reliably in a workplace.
Vals says its approach centers on economically useful tasks in law, finance, coding, health care and other domains. Its published methodology describes three layers: an open validation set that illustrates the type of work being tested, a larger private validation set available under license, and a test set that remains private and produces the public benchmark result.
The company also reports measures beyond accuracy, including latency, operating cost, tool use and qualitative failure patterns. Error bars are intended to show statistical uncertainty in the benchmark itself, though Vals notes that they do not capture every source of variability, including prompt choices, deployment settings and the stochastic behavior of language models.
Andreessen Horowitz, which led the funding round, said Vals keeps test sets private, limits how often they are run and can return results within hours of receiving model access. The investor’s announcement framed the company as a trust layer between model developers and users. That description is promotional, but the operational problem is real: a model that tops a general leaderboard may still be a poor fit for a specific workflow.
Analysis: Procurement needs task-level evidence
For enterprise buyers, the central change is that model selection can no longer be reduced to a vendor name or a single composite score. Legal research, financial analysis, software debugging and customer support place different demands on accuracy, context length, tool access, latency and cost. The right evaluation should resemble the work a company plans to deploy.
That means procurement teams should ask vendors and evaluators to define the task, scoring rubric, sample size and acceptable failure threshold before comparing results. They should also measure the full system rather than the underlying model alone. An agent’s performance depends on its prompts, retrieval layer, tools, permissions and surrounding software, all of which can change the outcome.
Private tests can reduce contamination and gaming, but they introduce a tradeoff. Buyers cannot independently inspect every item in a sealed test set, so they must assess the evaluator’s methodology, conflicts, security controls and process for retiring benchmarks that models have saturated. Reproducible procedures and clear uncertainty reporting become as important as the score.
Procurement teams should also require a baseline from their own environment. A benchmark built from representative documents, tools and approval steps can reveal whether a model’s published advantage survives contact with company data. The comparison should include human review, rollback conditions and a record of which model and system version produced each result.
The evaluator still has to earn independence
Vals is paid to test models and is backed by investors with broad interests across the AI market. Neither fact invalidates its results, but both make governance important. Customers need to know who funded an evaluation, who selected the tasks, whether a developer received advance access and what the evaluator can publish when findings are unfavorable.
Continuous evaluation will matter more as models and agent frameworks change after procurement. A useful program should rerun tests after model updates, monitor performance drift and include adversarial cases that reflect the buyer’s data and permission boundaries. Cost and speed should be evaluated alongside quality, because a slightly more accurate system may be operationally worse if it is too slow or expensive at scale.
The reported growth at Vals suggests that enterprises are beginning to pay for this discipline. The larger test is whether model evaluation develops into a credible independent function rather than another marketing layer. Buyers should welcome better benchmarks while insisting on evidence about how those benchmarks were built, governed and kept current.