What is a token in AI, and why it decides what you pay

Models such as GPT, Claude and Gemini do not see letters or words the way we do. Before reading a text they split it into tokens, and when they answer they produce one token after another. Common words are often a single token, while long or rare ones are split into several pieces.

The token is also the unit you pay for when you use the API, and the unit that measures how much text the model can hold in mind at once.

An example in an SMB

An online shop uses a model via the API to answer customer emails. Every email it reads is input tokens, every reply it writes is output tokens, and usually output tokens cost more. If the replies are longer and more polite than they need to be, the bill grows without the customer getting any benefit. Asking the model for short replies is the first saving.

The most common mistakes

• Comparing model prices by looking only at the cost per token, without counting how many tokens each model produces for the same job.
• Thinking a token is a word: to estimate costs on Italian text, it is better to measure on real examples.
• Forgetting that the fixed instructions and the documents attached to every request are paid for too, every time.

Why it matters to you

If you use a ChatGPT or Claude subscription for your employees, you never see the tokens. If instead a provider sells you a chatbot or an automation on a pay-as-you-go basis, the bill is made of tokens: ask how they measure them and how they keep them down.

All the AI terms explained simply are in the glossary.