Two scatter charts of the Artificial Analysis Intelligence Index for 37 configurations of 22 language models, including reasoning-effort variants of frontier models. The first plots intelligence against tasks completed per dollar: fifteen configurations form the cost frontier, from GPT-5.6 Luna at low effort near 114 tasks per dollar and intelligence 34, up to Claude Opus 5 at max effort with 0.43 tasks per dollar and intelligence 63. The second plots intelligence against tasks completed per second: fifteen configurations form the speed frontier, all of them OpenAI GPT-5.6 or Anthropic Claude configurations, from GPT-5.6 Luna without reasoning at about one task per 14 seconds up to Claude Opus 5 at max effort at one task per roughly 450 seconds. A third chart is a rotatable three-dimensional scatter of all three axes at once: twenty-three configurations are unbeaten across intelligence, cost and speed jointly, including GPT-5.6 Luna at xhigh effort and GPT-5.6 Terra at max effort, which appear on neither two-dimensional frontier. A full data table follows the charts.
LLM economics · artificialanalysis.ai · retrieved Aug 11, 2026
The Artificial Analysis Intelligence Index plotted against task throughput on two budgets: dollars, then seconds. Frontier models are shown at every benchmarked reasoning-effort level, connected by dashed effort curves. Up and to the right is better. The cobalt line marks each Pareto frontier: configurations nothing else beats on both axes at once.
Task progress per dollar — the inverse of cost per Intelligence Index task.
Task progress per second — the inverse of end-to-end time per Intelligence Index task. Same configurations, wall-clock budget instead of dollars. Muse Spark 1.2 publishes no time figure and is absent here.
All three at once — intelligence up, task progress per dollar and per second on the floor axes (both log). The translucent staircase is the frontier surface: every configuration on or below it is matched or beaten on all three axes at once by a cobalt point at a step corner. Wall shadows are the two 2-D frontiers above; drag to spin a full turn. Muse Spark 1.2 (no time data) is absent.
GPT-5.6 Luna’s effort settings span 21 to 114 tasks per dollar, and four of its six configurations sit on the cost frontier — at low effort it does an Intelligence Index task for under a penny. Only DeepSeek V4 Flash 0731 breaks the Luna run, dominating its xhigh setting.
Walking the cost frontier from intelligence 49.9 to 63.1 multiplies cost per task 86×. Effort dials show the same shape in miniature: Claude Opus 5 pays 5.5× from low to max for +10.6 points, and its final step (xhigh→max) pays 1.3× for just +0.5.
GPT-5.6 Sol (max) is edged off the cost frontier by Claude Opus 5 (high) — a higher index (61.5 vs 60.9) at essentially the same cost ($1.227 vs $1.231 per task). GLM-5.2 holds its frontier slot at $0.31 per task, the only open-weights model left in the middle of the line; Claude Fable 5 remains dominated only by its Opus 5 siblings.
Every point on the tasks-per-second frontier is an OpenAI GPT-5.6 or Anthropic Claude configuration — and Claude Fable 5, off the cost frontier, is on this one. The open-weights models that rule the cost chart fall away when the budget is wall-clock: Kimi K3 (max) takes ~9 minutes per task.
GPT-5.6 Luna (xhigh) and GPT-5.6 Terra (max) sit on neither 2-D frontier — each loses on pure cost and on pure speed — yet nothing beats them on all three axes at once. Jointly, 22 of the 36 timed configurations are unbeaten.
Every benchmarked configuration, including reasoning-effort variants of frontier models, ranked by intelligence index (ties broken by task progress per dollar). Hover a row to locate its point on the charts.
| # | Model | Lab | Intelligence | Cost / task | Tasks / $1 | Time / task | Tasks / sec |
|---|