The company says the model was trained in multi-user simulations with evolving goals and preferences over long interactions, rather than optimizing only for task completion or shallow preferences.
Humans& reports a 59.1% score on the Collaborative Intelligence Benchmark at high effort and 39.0% cooperative success on CooperBench. It also reports that Ampersand can incur higher interaction costs than other evaluated models, including by offering coaching when users would prefer a direct answer or to work things out themselves. These are company-reported results; the supplied source does not establish independent validation or real-world deployment outcomes.