← API pricing · glossary

LLM API pricing glossary

The words on an LLM API price sheet, defined once. Each one links to the page where ArtificialWatch shows that number for every live model.

Token · Input tokens · Output tokens · Reasoning (thinking) tokens · Price per million tokens · Cached input (prompt caching) · Cache write · Long-prompt rate · Context window · Max output · Batch discount · First-party price

Token

The unit an LLM reads and writes: a word, part of a word or a punctuation mark. In English a token averages about four characters, so 1,000 tokens is roughly 750 words; each model's tokenizer counts a little differently. The cost calculator estimates tokens from pasted text.

Input tokens

Everything sent to the model in one request: the system prompt, the conversation so far, documents and tool definitions. Billed at the input price, usually the lower of the two rates.

Output tokens

What the model writes back. Billed at the output price, usually several times the input price, so long replies drive cost. See every model's two rates on LLM API pricing.

Reasoning (thinking) tokens

The working a reasoning model writes before its answer. Most makers bill it as output tokens even when it is hidden from the reply, so a short answer can still cost a lot. Count it in the output figure when you price a workload.

Price per million tokens

The standard unit for LLM API prices: dollars per 1,000,000 tokens, quoted separately for input and output. A request's cost is its tokens times the rate, divided by one million; the calculator does the multiplication for every model at once.

Cached input (prompt caching)

Input the API has seen recently and kept, such as a long system prompt that starts every request. Reading it from the cache is billed at a lower cached-input price; how much lower varies a lot by model. Compared on prompt caching prices.

Cache write

Some makers charge to store a prompt in the cache the first time, and set how long it stays. The saving comes on later requests that reuse it, so caching pays off when the same prefix repeats within that window.

Long-prompt rate

A higher price some makers charge once a prompt passes a size threshold; past it, the whole request is billed at the higher rate. Each model's threshold is on the context window comparison.

Context window

The most tokens a model handles in one request, the prompt and the reply together. A long document, a codebase or a long chat needs a big window. Ranked on LLM context windows.

Max output

The longest reply a model can write in one request. It is separate from, and usually much smaller than, the context window.

Batch discount

Some makers offer a lower price for jobs that can wait hours for their results. ArtificialWatch lists the standard rates, not batch rates.

First-party price

What a model's own maker charges on its own API. Clouds and resellers that also serve the model may charge differently; ArtificialWatch quotes first-party list prices only, by lab on pages such as Anthropic API pricing.