What AI models cost to use

How the index is calculated

The Ensurva AI Price Index answers one question: what does it cost, at list price, to use a typical current AI model through its API? It is quoted in US dollars per million tokens.

Each model gets one price, its blended price: 70% of its input rate plus 30% of its output rate. Most production work reads far more than it writes, and a fixed weighting keeps every model on the same footing. Cached input, batch and priority rates are left out.

The index is the geometric mean of the blended prices of every model in the basket. Each model counts equally, and a change of a given percentage moves the index by the same amount whether it happens to a cheap model or a dear one. An ordinary average would be dominated by the most expensive few: today it would read US$4.17 against the index's US$1.79.

The history is worked out from the same price records for every Friday since the basket began, so correcting a past price corrects every week it touched. A new model joins at its own price, and a model that leaves takes its price with it, so the index tracks what the market charges now rather than any one model's price over time.

What it does not measure

Published list prices, not what any organisation pays. Real bills are usually lower: negotiated discounts, shorter prompts, caching, routing simple work to cheaper models, batch processing and cloud credits all reduce them. The savings calculator models those levers.

Tokens are not the same size everywhere. Each provider splits text into tokens its own way, and the same English text can take noticeably more tokens on one model than another. The index does not adjust for this, so per-token prices slightly flatter models whose tokeniser produces more tokens.

Some providers charge more for very long prompts or at peak hours. The index uses each model's standard rate for its base context band.