What is an AI token and why is it the unit that defines your cost

If you've ever looked at an AI API bill or the pricing table for ChatGPT, Claude, or Gemini, you've seen the same phrase repeated: per million tokens. The token is the unit behind almost everything you pay for when using artificial intelligence. Understanding it is the first step to controlling costs instead of getting surprises at the end of the month.

This article explains what a token is, how it's counted, and why this seemingly technical unit directly determines how much you spend.

So, what is a token?

A token is a piece of text that the model processes at a time. It's not a word, nor a letter; it's something in between: a subset of characters that the model uses internally to read and generate language.

In practice, the approximate rule is simple: one token equals about four characters in English, or about 0.75 words. Put another way, one million tokens correspond to about 750,000 words. A short word can be a single token; a long or rare word can be broken into two or three.

Why is text divided this way?

Language models don't read whole words like we do. They work with a fixed vocabulary of tokens and convert all text into this internal currency before processing it. This division allows them to handle any word, including proper nouns, code, or terms they've never seen, by assembling them from known pieces.

That's why token count varies with language and text type. Portuguese tends to generate more tokens than English for the same idea, and text with many symbols, numbers, or code is usually heavier in tokens than simple prose.

Input and output: two types of tokens

This is the most important distinction for your wallet. When you use an AI API, you pay for two types of tokens:

1. Input tokens. Everything you send to the model: your query, system instructions, and any context you add, such as documents or conversation history.
2. Output tokens. Everything the model generates in response.

And here's the detail that catches many people off guard: output almost always costs more than input. In most current models, the ratio is about five to one. That is, generating text is more expensive than sending it.

Why this defines your cost

The bill for an AI API is, in essence, a simple multiplication:

(input tokens × input price) + (output tokens × output price)

Everything that increases the token count increases the bill. Long prompts, conversation histories that grow with each turn, large documents sent as context, extensive responses. In a single call, the value is tiny, but an application with thousands of users processes hundreds of millions of tokens per month, and there, every token counts.

That's why teams running AI in production constantly talk about tokens. It's not technical obsession; it's cost management. Those who understand the unit understand where they are spending and where they can save.

What this means in practice

Three useful conclusions to draw from this:

• Shorter, more direct prompts cost less. Every extra word in the input is money.
• Longer responses cost more, and they cost double. Limiting output size, when it makes sense, has a direct impact.
• The choice of model is the biggest factor. The price per token varies enormously between models, and using the right model for each task is the most powerful lever of all.

Conclusion

The token is the currency of artificial intelligence. Everything you send and everything you receive is converted into this unit, and it's what you pay for. Understanding what a token is, the difference between input and output, and how the count grows is the foundation for using AI sustainably, without costs becoming a surprise.

Next, it's worth understanding how prices compare between providers and what strategies exist to reduce token consumption without losing quality.