Artificial Analysis updates Intelligence Index to prioritize agentic workloads and granular efficiency metrics
Artificial Analysis has updated its Intelligence Index to version 4.1, incorporating new benchmarks for agentic workloads and granular metrics for cost, time, and token efficiency.
1. Updates to the Intelligence Index
Artificial Analysis has released version 4.1 of its Intelligence Index, shifting its focus toward agentic workloads. This update includes the upgrade of several key benchmarks: Terminal-Bench Hard has been updated to 2.1, τ²-Bench Telecom has been replaced by τ³-Bench Banking, and GDPval-AA has been upgraded to v2. The latter now features a higher turn limit of 250 for longer-horizon trajectories and a rotating panel of frontier-model judges. Additionally, the IFBench benchmark was removed due to saturation, as it no longer effectively distinguishes between top-tier models.
2. New Performance Metrics
The v4.1 update introduces three new per-task metrics to provide a more granular view of model efficiency: cost per task, time per task, and tokens per task. These metrics are calculated by dividing the total cost, time, and output tokens required to complete the full Intelligence Index by the number of tasks involved. Furthermore, the index now accounts for cached input tokens to better reflect the real-world expenses associated with running these models.
3. Current Model Rankings and Efficiency
As of June 15, 2026, Claude Fable 5 holds the highest score on the Intelligence Index, though it is currently unavailable. Among accessible models, Claude Opus 4.8 leads with a score of 56, followed by GPT-5.5 at 55. In the open-weights category, DeepSeek V4 Pro and MiniMax M3 lead with scores of 44. Regarding efficiency, DeepSeek V4 Pro is noted for its cost-effectiveness at $0.04 per task, significantly lower than proprietary alternatives. In terms of speed, Gemini 3.1 Pro Preview stands out by completing tasks in 1.6 minutes, while other models range from 1.5 to 13.5 minutes per task.