All reasoning levels from low to max. Top left is better: smarter and cheaper or faster. GPT-5.6 Sol and Claude Opus 5 are hidden by default.
Weighted average cost (USD) per Intelligence Index task, at API prices
Codex Pro 20x (GPT models): user measurements with Sol come to about $22–24 of API value per percent of weekly usage, i.e. $2,200–2,600 per full week (Devforth cites $2,560); default $2,400. With Astra, one percent buys fewer API dollars: about $1,500–1,800 per week; default $1,650. There is no separate measurement for Luna and GPT-5.6 Sol, so they use the Sol budget. Codex credits are exactly proportional to API prices (1 credit ≈ $0.04), but the weekly budget in credits differs per model.
Claude Max 20x: about $1,300 of API value per week for Opus. Fable may only use part of the weekly limit (default 50%), so Fable uses $650 per week.
Percentage = API cost of the task ÷ that model's weekly budget. Adjust the amounts if your experience differs.
Average time to complete one Intelligence Index task; left is faster
Source: artificialanalysis.ai, retrieved 22 Sep 2026 (time per task at about 21:45 CEST; Artificial Analysis updates these measurements). Claude levels = Adaptive Reasoning with low to max effort (max = Max Effort, Default Fallback). No time per task has been measured yet for Claude Opus 5.5 max. The Pareto line only covers the visible models.