Understanding token usage

What a token is

Models read and write text in tokens, not characters or words. A token is roughly a few characters of English — a short word is one token, a long or unusual word is several. Both what you send (input tokens) and what the model returns (output tokens) are counted.

Input and output are priced separately

Every model has two rates: one for input tokens and a higher one for output tokens. Output is the more expensive side, so a short question that produces a long answer is dominated by the output it generated.

The exact per-million-token rate for each model is shown on the pricing page and is the source of truth.

Why a short prompt can cost more

The text you type is rarely the whole request. When you use a coding agent like Claude Code, each turn also sends context the model needs to answer well:

• Conversation history — earlier messages in the session are resent so the model remembers them.
• Project and file context — the tool includes relevant files, directory listings, or tool output.
• System instructions — the tool’s own guidance to the model travels with every request.

So a three-word follow-up can carry thousands of input tokens of accumulated context. This is normal for agentic tools and is why input token counts grow over a long session. It is not an extra fee — you are billed for the tokens the model actually processed.

Keeping usage down

• Start a fresh session for an unrelated task so old history is not resent.
• Ask for focused output when you don’t need a long explanation.
• Reuse stable context so it can be cached — see prompt caching.