Reading the unit
Vendors quote rates per million tokens because a single token costs a tiny fraction of a cent, which is awkward to print and compare. A million tokens is roughly 750,000 words of English by OpenAI's rule of thumb — several novels — so the per-million figure is a convenient unit, not a typical request.
Every rate card lists at least two numbers, one per direction. Input covers everything sent to the model; output covers everything it writes, including hidden reasoning on thinking models. Most cards add a third for cached input, and some add cache writes, a batch column, flex or fast columns and a long-context card. Each column is its own meter on the bill.
In our indexPer-million-token rates in our index
Counted at build from our price index: standard (uncached, real-time) rates in US dollars per million tokens, for the 163 paid models, among the 167 current public models we price, from 33 vendors. Each rate comes from the vendor's own price page, read between 2026-07-11 and 2026-10-09. How we collect prices.
From a rate to a cost per request
The arithmetic is the same for every meter: tokens × rate ÷ 1,000,000. A request with 3,000 input tokens and 500 output tokens on Claude Sonnet 5.5 — $2 in, $10 out per million — costs 3,000 × the input rate plus 500 × the output rate, each divided by a million. Multiply by monthly requests for the monthly bill. The worked example below does exactly that at build time, on today's list rates.
Two things make the real number differ from the list rate. Cached input and batch discounts lower it. Tool fees, search fees and multimodal inputs raise it, because they are billed in other units: Anthropic charges for web search per 1,000 searches on top of tokens, and xAI bills its server-side tools per 1,000 calls.
Worked exampleOne rate card, one workload, one bill
30,000 requests a month on Claude Sonnet 5.5, priced at its current list rates, checked on 2026-10-09.
Thirty thousand requests a month at 3,000 tokens in and 500 out. The input line and the output line are each tokens × rate ÷ one million; the blended line divides the total by every token billed.
Change the workload or the model in the LLM API cost calculator.
Comparing models on the rate
A lower price per million does not guarantee a cheaper task. OpenAI's own token guide warns that models tokenize the same text differently and write different amounts of output and reasoning, so the total can come out the other way. Compare on a blended rate that matches your traffic, or better, price the actual workload.
Currency is the other trap. Several Chinese vendors publish a yuan sheet for their home market and a separate dollar sheet for international customers, and the two are not linked by an exchange rate. Our index takes the dollar sheet where one exists and, where only a yuan price is published, converts it at one dated exchange rate kept with the snapshot.