The index combines four evaluations: AA-Omniscience-Decision for knowledge and hallucination, HLE-MultiChoice for reasoning, When2Call for tool calling, and WildGuardTest for classification. Each contributes 25% of the weighted score.
The evaluation page also presents calibration, median task time, and estimated cost per million tasks. Artificial Analysis says it conducts the evaluations independently. The results offer a comparison framework for bounded decision tasks, but the supplied evidence does not show that index performance predicts outcomes in any particular production workflow.