What is a token in AI?
So what is a token in AI? It is the piece of text a model counts with: not a letter and not exactly a word, but something in between. A common chunk of characters the model turns into numbers before it can work with them.
What is a token in AI, from the model's side
Models do not read words, they read tokens: frequent fragments of text mapped to numbers. In English a token averages about three quarters of a word, so a hundred words is roughly a hundred and thirty tokens.
Languages with heavy inflection fare worse. The same sentence in Hungarian or Finnish can cost nearly twice the tokens it costs in English, which is a property of the model's vocabulary and nothing you can write your way out of.
Why it decides your bill
APIs charge per token, and not just for your question: the material you pasted, the conversation so far and the model's own answer all count.
That is why a long conversation gets more expensive with every turn. At the tenth question you are paying for the previous nine as well, because the whole history goes back through the model each time.
How to spend fewer
When the subject changes, start a new conversation. The model will not take offence, and you stop dragging fifty irrelevant paragraphs along.
With a large document, give the relevant part rather than all of it. Long context is not only more expensive, it is also less accurate, because the point gets lost in the filler.
Common questions
What is a token in AI, in one sentence?
It is the unit of text a model reads and charges for. Everything you type and everything it answers is counted in tokens, and so are the limits.
How many words is a token?
In English, about three quarters of a word on average. Punctuation and rare words split into several tokens, so code and names cost more than plain prose.
Why does a long conversation cost more?
Because the entire history is resent with every new question. The cost of a conversation grows faster than the number of messages in it.
Do tokens limit how much I can paste?
Yes. The context window is measured in tokens, so a document that exceeds it has to be split or summarised before the model can work on it.