Each point is one model variant, usually one reasoning effort level. A line joins the variants of one model family. Point at a variant, or tab to it, to see its details. Click it to pin it, then point at another variant to compare the two. Right-click a variant, a label or a model row for more actions.
On a touch screen, tap near a variant to pin it and see its details, and tap a legend entry to highlight its family. Long-press a variant, legend entry or model row for more actions.
The X axis shows the cost to run one task of the Artificial Analysis Intelligence Index test suite. On the log scale, equal distances stand for equal cost ratios: each tenfold step in cost takes the same width. The Y axis shows the benchmark picked in the header. A higher score is better on all of them.
The dashed line is the Pareto line. A variant is on it when it scores higher than every visible variant that costs less, and no variant at the same cost scores higher. When variants tie on both cost and score, one of them is on the line. These variants give the best value for money. Zoom Y to Pareto fits the score axis to them.
Only models with a score for the picked benchmark appear. Whatever the benchmark, the cost axis uses the Intelligence Index task costs and the response time filter uses the Artificial Analysis response times.
Deprecated models are hidden unless Show deprecated is on, so some benchmarks, such as Math (AIME 2025), show only a few points by default.
The Claude plan prices Anthropic models, and the Codex plan prices OpenAI models that are not open weights. With a plan picked, the page scales the API cost per task by the plan's monthly fee divided by the API value of its allowance. This assumes you use the whole allowance over a long window. It is a comparison estimate, not a provider invoice, and real limits and usage differ.
Reasoning models can think for minutes before they answer. The slider hides variants whose median end-to-end response time is above the limit. That is the time from the request to the last token of a 500-token answer. Hidden variants stay on the chart as faint dots, and the Pareto line is computed again from the variants that remain. Variants with no measured response time are never hidden.