Rehearsals, a consumer research platform that simulates AI twins of real people, has launched the Human Behavior Bench in partnership with Google DeepMind and Kaggle. The new resource is described as the first public benchmark designed to measure how well AI models predict and understand human preference and action. Rehearsals also previewed its internal model, Rehearsals Engram 1.0, which topped the initial leaderboard with a distribution agreement score of 94.1, ahead of frontier systems from Google, Anthropic, OpenAI, and SpaceXAI.
Benchmark Design and Scope
The first version of the benchmark draws on interviews with 1,244 U.S. adults across 750 categories, producing 95,000 responses. An 80% private holdout group maintains the validity of the evaluation, while the remaining portion is public on Kaggle so that other organizations can score their latest models. Rehearsals and DeepMind designed the benchmark to evaluate observed human decision making and stated reasoning across health, habits, decision making, and values.
Model Performance and Key Findings
Rehearsals Engram 1.0 achieved a distribution agreement score of 94.1, followed by Gemini 4 Argon at 80.6 and GPT-5.6 Sol at 79.5. Claude Opus 5.5 scored 79.3, Gemini 3.8 Flash scored 79.1, GPT-6 Astra scored 79.0, and Grok 4.6 scored 76.2. According to the company, the gaps show that state of the art AI models still miss critical categories of human understanding, especially around health, habits, decision making, and values.
Misconceptions About Human Behavior
The benchmark exposed several incorrect assumptions made by current AI systems. For example, AI models think most people over 65 do not watch short form videos, but 80% of respondents reported watching them in the previous week. Other gaps include AI estimates that 30% of parents with toddlers cook, while 66% actually do, and that only 5% of older Midwest shoppers pay by phone, compared with a reported 25%.
Leadership Perspective
Liam Bolling, CEO and co-founder of Rehearsals, said the industry focus of 2026 was on agents that complete long running tasks and master software engineering. He explained that the company's bet for 2027 and beyond is a renewed focus on values and cultural alignment to humanity as AI becomes more personal, which is why Rehearsals co-created the benchmark with Google DeepMind. Bolling added that no public benchmark previously determined a model's understanding of humanity when new systems launched.
Future Expansion and Company Background
Rehearsals expects the next version of the benchmark to launch before the end of 2026, with more participants, four additional countries, and a broader spectrum of focus areas. The company operates an AI consumer research platform that simulates AI twins of real people for pricing, messaging, ad creative, and product concept tests in about 30 minutes. Rehearsals is backed by Andreessen Horowitz Speedrun, 25madison, Plug and Play Ventures, and Factorial Fund, with angel investors from OpenAI, Google DeepMind, and Scale AI.
The Human Behavior Bench provides a public scorecard for evaluating how well AI models align with real human preferences, moving beyond coding and task completion metrics. Rehearsals Engram 1.0 currently leads the leaderboard and powers the predictions that the company's customers rely on for business decisions across pricing, messaging, and product concepts. More results, findings, and updates are available on the interactive benchmark site and on Kaggle, where any organization can score its latest model.