Last updated: 2026-06-22

Token cost explained

The token cost of a large language model API call is the price charged for a single request, computed from the number of tokens sent to and received from the model and the model's published per-token rates. Providers meter usage in tokens — short fragments of text, roughly four characters of English each — and bill the prompt (input) and the completion (output) separately, because the two are charged at different rates.

How pricing is quoted

A single token costs a minuscule fraction of a currency unit, so providers publish rates per one million tokens to keep the headline number readable. A rate of $3 per million input tokens means each input token costs three millionths of a dollar; an output rate of $15 per million is five times as much. Output is the more expensive side because generation produces one token at a time, each requiring a full pass over the model, whereas the input is processed in a single parallel pass.

The formula

With $n_{in}$ input tokens at price $p_{in}$ per million and $n_{out}$ output tokens at price $p_{out}$ per million, the two billed components are

and the call cost is their sum:

The division by $10^6$ converts the per-million quote into a per-token rate before multiplying by the token counts.

Worked example

Consider a request with $n_{in} = 1500$ input tokens and $n_{out} = 500$ output tokens against a mid-tier frontier model priced at $3 per million input and $15 per million output. The input portion is

and the output portion is

giving a total of C = 0.0045 + 0.0075 = \0.012$ for the call. Although the prompt is three times longer than the completion, the two halves contribute roughly equally to the bill, because the output rate is five times the input rate. This is the defining feature of token pricing: a short completion at a high rate can cost as much as a long prompt at a low rate.

Notes and variations

The model assumes a flat per-token rate. In practice the effective rate can fall below the headline figure: prompt caching reuses repeated context at a discount, and asynchronous batch processing trades latency for a lower rate. Token counts also depend on the tokenizer and the content — code and non-English scripts tokenize less efficiently than plain English, producing more tokens per character.

Application

Per-call cost is the atomic unit from which larger estimates are built. Multiplying it by request volume gives recurring spend, the subject of the LLM Monthly Cost Calculator, while an agentic workflow that issues many calls per task is handled by the Agent Cost Calculator. To work in the opposite direction — how many tokens a fixed budget buys — see Tokens per Dollar Calculator.