What is a token in artificial intelligence?

By Gianluca Bonomo — Technology Trainer and Instructor

Published on October 5, 2026

Have you ever seen «token limit» or «price per token» and pretended to understand? No worries: the word sounds technical, but the idea is very simple.

In this lesson you will find out what a token is, how artificial intelligence uses tokens to read your question and to write its answer one small piece at a time, and why it is worth knowing. There are four examples where you watch an answer being born token by token.

1. The problem: AI does not read like us

You see words and sentences. Inside, the computer sees only numbers.

When you write «Hello, how are you?» to a chatbot, nothing inside the machine «understands» those letters the way you do. First the text has to be taken apart into small pieces, and each piece turned into a number. Only then can the model work with it.

Those pieces are called tokens. Think of toy building bricks: with just a few kinds of brick you can build castles, spaceships and even a platypus. With a limited set of tokens, AI builds any sentence.

2. What is a token?

A token is not always a word. It can be a whole word, part of a word, a punctuation mark, or a space attached to the word that follows. Common words, which AI has seen millions of times, are usually a single token. Rare words are split into several pieces.

2.1 How many tokens is a text?

A handy rule of thumb, especially for English: one token is about four characters, and one hundred tokens are about seventy-five words. Languages with long words and many endings, like Italian, usually use somewhat more tokens than the same text in English.

2.2 From text to numbers

AI has a closed list of tokens it knows, the vocabulary: tens of thousands of pieces, each with a code number. The text you write is cut into pieces from that list, and each piece is replaced by its number. That series of numbers is what goes into the model.

The same thing happens in reverse: the model produces a number, and the number is translated back into the piece of text you see appear on the screen.

3. How AI writes: one token at a time

AI does not write the answer all at once: it builds it one piece at a time, like when you compose a message on your phone.

The mechanism always repeats the same way. The model reads all the text so far, calculates how likely each possible token is to come next, picks one and adds it. Then it starts again, with one more token to read.

3.1 Why the same question gives different answers

The model does not always pick the most likely token: sometimes it also picks among less likely ones. A setting called temperature controls how adventurous it is. Low: predictable, uniform answers. High: more imagination and more surprises. That is why asking the same thing twice gives you two different texts.

4. Four answers built token by token

Let us see it live. Each example has a question and the answer that is born one token at a time. Read the tables from top to bottom: each row adds a piece, and the last column shows the text that exists at that point. The splits are illustrative, but the mechanism is the real one.

4.1 Example 1: an easy question

At step 6 the model chooses among many candidates, each with its own probability. This is how that moment might look:

4.2 Example 2: a piece of advice

4.3 Example 3: the rare word

4.4 Example 4: the silliest excuse

The funniest choice is at step 2. When the request is creative, no candidate dominates and temperature makes the difference:

5. Funny tokens: the museum of pieces

If you think tokens are boring, you have not yet seen what happens when AI has to cut up certain words.

The longer, rarer or weirder a word is, the more pieces are needed to rebuild it. Here is our little museum: the cuts are examples, but the principle is real.

• Mile-long words. «Supercalifragilisticexpialidocious» is a ten-piece puzzle: AI assembles it brick by brick.
• Capital letters. «HELLO» and «hello» are two different things to AI: the shouted version can cost more tokens.
• Stretched letters. The more «y» you add to «Heyyyyy», the more pieces it needs: stretching letters stretches the bill.
• Numbers. Even a year like 2026 can be split in two, for example «20» and «26».
• Your name. «Gianluca» might become «G», «ian», «luca»: try it in the counter and see how many pieces yours is cut into.

6. Why it is worth knowing

Tokens are not an engineer's curiosity: they explain three things you meet every day when you use a chatbot.

6.1 Cost

Many AI services, especially those for businesses and developers, are paid by the token: both the ones you send and the ones you receive are counted. Short prompts and focused answers cost less. Prices change often: always check the service's current price list.

6.2 Limits and memory

The model can only take into account a limited amount of text, the context window, measured in tokens. Your question, the conversation so far and the answer all fit inside it. When the conversation gets too long, the oldest parts fall out and the model «forgets» them.

6.3 Funny mistakes

The model sees pieces, not letters. That is why it can stumble on tasks that are trivial for us, like counting how many times a letter appears in a word or spelling a word backwards. It is not stupid: it is looking at the bricks, not at the individual grains of plastic.

A token is a brick. AI has never written a whole sentence in one go: it has always stacked one piece on top of another.

— Gianluca Bonomo

Hands-on exercise

Fifteen minutes, a browser and your favourite chatbot. Nothing to install.

1. Write a sentence you like, for example the first sentence of a work email.
2. Search online for the token counter («tokenizer») offered by the provider of the chatbot you use, and paste your sentence in.
3. Note how many tokens and how many characters there are, and look at how it was split.
4. Try three variants: the same sentence in another language, a long rare word, and a line of numbers only. Compare the counts.
5. Funny token contest: compare «hello», «HELLO», «helloooooo» and your name. Which costs more?
6. Final game: on your phone keyboard type the start of a sentence and always tap the first suggestion. After ten words read what came out: you have just done, by hand, what AI does.

Questions that come up in class

Glossary

Further reading

• AI providers' documentation on tokens and token counting — official service pages: rules of thumb (about four characters per token) and online counters
• Sennrich, Haddow, Birch - Neural Machine Translation of Rare Words with Subword Units (2016) — the paper that popularised splitting text into subword pieces
• Vaswani et al. - Attention Is All You Need (2017) — the architecture behind modern language models
• Gianluca Bonomo, Niente paura, è solo IA — the reference book (in Italian)