Methodology version 2.0. List date Oct 13, 2026. Base 100.
This index
Capability = Epoch AI's Capabilities Index (ECI), which combines a model's results across many benchmarks into one score. Scores by Epoch AI, used under CC BY 4.0.
Price = each model's published API price, in US dollars per million tokens, blended as (3 × input price + output price) ÷ 4, as in the AI inference cost index; a model's price for a month is its median that month. Free models and variants of a model are excluded.
Method = a monthly chained hedonic index, the method statistics agencies use for products whose quality keeps rising. Each month, the logarithm of every scored model's price is regressed on its capability, with a dummy for each month, over that month and the two before it. The month's dummy less the previous month's is the change in price at equal capability; performance per dollar moves by its inverse, chained from 100 in the first month.
So a price cut and a better model at the same price both raise the index, and every model on sale counts, not only the cheapest. A price far from the fit (more than three robust standard deviations, such as a promotional or mistyped price) is left out of the month's second fit. A month with fewer than five scored models holds the index.
Before 2025 copies of the price list were taken weeks apart, so a month without one repeats the month before, and its price changes appear in the next month with a copy.
A model's score counts from its release date, so a model name later reused for a newer model never lends that model's score to the past.
Changes
No changes since the list date.
Sources: Epoch AI's benchmark data (CC BY 4.0), and AI models' published API prices with archived copies of them before the list date.
Epoch refits its index as models are added; each day uses the scores published on or before it. Values before the list date use scores published later and are hypothetical.
How the model performance per dollar is built · Indices · The Inference · The Inference