What AI models cost, per million words.
API prices from Anthropic, OpenAI, Google, xAI and DeepSeek, read from their own pricing pages every hour and converted to one unit. Each model splits text into tokens its own way, so a token on one model is not a token on another. Words are the same everywhere.
51 models from 5 makers. Every pricing page was last read at 5:39 PM Eastern, Oct 7, 2026 or later. No price changes this week.
20 of 51 models. Three words in for every word out, as in a chat.
Estimated: from the same maker’s measured models, for a model we cannot reach to measure. In Brass: one of the eight models in the Brass chat.
How the board works
- Every price is the maker’s own standard API price per million tokens, read from its pricing page every hour; each row links to that page. Batch discounts, long-context tiers, regional uplifts and cached input are left out of the headline figure; each model’s page lists its cached-input price.
- Words come from one test: we send the same 457-word English passage to each model and count the tokens it bills. A model that needs fewer tokens for the same words costs less than its token price suggests.
- When a maker announces a price change with a date, the board shows the new price under the current one, and when a maker notes a promotion, the board says so beside the price.
- A changed price is shown once a second read, half an hour or more after the first, agrees. If a page has not been read for two hours, the board says when it last was; after a day, that maker’s prices are marked out of date.
- More makers join the board as we build readers for their pricing pages.
Use the data
market.csv and market.json, updated every hour. The prices are the makers’ published facts, each with the page it came from. Brass’s own figures, the tokens per thousand words and the prices per million words, are free to reuse under CC BY 4.0; credit “Brass market, brass.adorellc.pro”.