What is a token? How LLM APIs bill you

A token is the smallest unit a large language model reads and writes, and the unit LLM APIs bill by. In English one token is about 4 characters or 0.75 words; for Chinese it depends on the model, with common characters around one token each. Prices are usually quoted per million tokens, with input and output priced separately.

Tokens versus words

Models do not read characters directly. A tokenizer first splits text into tokens: common words become one token, rare words are split into several pieces. Token counts track text length, but there is no fixed ratio:

The same text produces different token counts on different models because each vendor has its own tokenizer. Use the table for a first budget, then check real usage numbers once you are live.

Input tokens and output tokens

A request's cost has two parts:

• Input tokens: everything you send — the system prompt, chat history, retrieved documents, tool (function) definitions and images converted to tokens.
• Output tokens: everything the model generates. A reasoning model's thinking counts as output too, so output can be large even when the final answer is short.

Output usually costs several times as much as input. Live prices for a few common models in their cheapest groups:

What one request costs

For token-billed models, the cost of a request is:

cost = input tokens ÷ 1,000,000 × input price + output tokens ÷ 1,000,000 × output price

Take a support chat turn: about 2,000 tokens of system prompt and history, and a 500-token answer. With deepseek-v4.1-flash:

The same turn 100,000 times a month:

For another model, swap in its prices. Every model's price in every group is on Models and pricing.

Context length is counted in tokens too

Each model has a context limit: the maximum number of input plus output tokens in one request. Go over it and the request fails or the answer is cut off (finish_reason is length). In long conversations the history keeps taking up input tokens, which costs money and crowds the context, so trim it or summarize it regularly.

How to see real usage

• Every response returns this request's input and output tokens in usage; in Chat Completions they are prompt_tokens and completion_tokens.
• The Usage page in the console lists model, tokens and cost for every request, and you can export it.
• To estimate before sending, use the vendor's tokenizer tool — but usage is what you are billed on.

Ways to use fewer tokens

1. Trim the system prompt: remove repeated instructions and examples.
2. Keep only the last few turns of a conversation and summarize the rest.
3. Cap answer length with max_tokens or max_completion_tokens.
4. Put long, unchanging prefixes (system prompt, product docs) behind context caching so the cached part is billed at a lower price — see What context caching is and how it saves money.
5. Use a small, cheap model for simple tasks and a large one only where needed — see How to choose a model for your use case.