also tokens, tokenization
The unit a model reads and writes, usually a piece of a word a few characters long. Prices, context limits and output limits are all counted in tokens, so the same text costs and fits differently across models and languages.
You have done this if
A prompt that fitted comfortably in English overflowed the context limit once the documents were in another language, or the bill grew because a long system prompt went out on every call.
Say it in a review
We budget in tokens, not characters: the system prompt and retrieved context are capped so the answer always has room.
On the AI Application map Model, LLM Gateway
Read Tokenization quietly sets your cost, limits and reliability