Halluminate works in two steps. It builds benchmarks to find where frontier models fail on real financial workflows, then turns those failures into RL environments that labs can train on. Its flagship benchmark, Westworld Finance Diligence Bench, was released on August 11, 2026. The company describes it as a benchmark of 88 problems that evaluates AI agents on a complete company acquisition due-diligence process, drawn from anonymized data from real private transactions. Fortune reported, as of October 1, 2026, that the highest average score was 51%.[1],[5],[2],[4]
Halluminate's research page dates its earlier releases. Web Bench, a benchmark of about 2,454 tasks across 452 real-world websites for browser agents, came out on May 30, 2025. Westworld v1, a set of sandbox environments on simulated websites for computer-use agents, followed on November 23, 2025. DealTrace Bench, a staged review of a real private-equity deal in five graded stages, came out on May 15, 2026. The sequence shows a move from general browser agents toward finance work.[5],[4],[2]
For now its customers are frontier labs rather than enterprises. Wu told Fortune that four of the top five closed-source U.S. AI labs are paying customers, and that the company is deliberately concentrating on that small group. He also describes a "Moore's law of environments": every six to eight months, the complexity of the company's environments needs to roughly double to keep pushing frontier models.[2]