The Artificial Analysis Intelligence Index v4.1 evaluates models across 9 different metrics, including GDPval-AA v2, 𝜏³-Banking, Terminal-Bench v2.1, Sci Code, Humanity's Last Exam, GPQA Diamond, Crit Pt, AA-Omniscience, and AA-LCR. The index methodology details each evaluation and its execution.

New benchmarks like AA-Briefcase Elo assess agentic knowledge work by aggregating rubric pass rate, analytical quality Elo, and presentation Elo. AA-Omniscience Index specifically measures knowledge reliability and hallucination, rewarding correct answers while penalizing inaccuracies and offering no penalty for refusal to answer. Scores range from -100 to 100.

Cost analysis compares the weighted average cost per Artificial Analysis Intelligence Index task, considering input, cache, reasoning, and answer tokens. Token usage is also measured by the weighted average output tokens per task. Pricing details include cache hit, input, and output prices per million tokens. Context window size, measured in tokens, is crucial for RAG workflows. Performance is further gauged by output speed in tokens per second and weighted average decode time per task, with lower times indicating better performance.