What it measures
The lowest price at which a given level of machine intelligence can be bought, tracked day by day. For each of four benchmarks Epoch AI runs (GPQA Diamond, OTIS Mock AIME, FrontierMath Tiers 1-3 and Chess Puzzles), and for each score level that models have reached (25%, 50% and 75% of the way from guessing to perfect), the index takes the cheapest cost per task of any model released to date that reaches that score. Eleven benchmark-and-level series in all, equally weighted.
Why these components
Raw token price is the wrong unit and moves the wrong way: frontier tokens have become more expensive as reasoning capability rose, and reasoning models spend many more tokens per answer. Cost compression is only visible when capability is held constant and the tokens a model actually spends are counted. Epoch's data does both: it prices each model's real benchmark runs, across reasoning settings and token budgets. Four benchmarks in three domains (science questions, competition and research math, games) and three score levels are used because compression arrives at different times for different tasks and capability levels: a score cheapens fastest soon after models first reach it.
How to read it
Higher readings mean a fixed level of capability costs less than it did across the trailing twelve months. Each component is read against its own past year, so the index rises when a new model sets a record low and eases back as that record ages. The dollar figures are the more legible number and are published as the component signals. Epoch's own estimate from the same data is that the cost of a given level of performance has fallen about 47% a quarter since 2023.
What it does not say
It says nothing about total AI spend, which has risen sharply even as unit costs collapsed. It is a list-price measure: it leaves out negotiated enterprise pricing, batch and cached-input discounts, and most self-hosted economics. It covers only the models and settings Epoch has tested, and costs under tighter budgets are estimated from each run's transcript. Four benchmarks are not all of work, and a benchmark score is not workplace performance.