What is a token in AI, and why are you charged per token?
Tokens are the small chunks of text AI models read and write. Learn how they work, how to estimate them, and why AI companies charge and set limits by token.
First published · AI Digested
A token is the unit of text an AI model actually works with. Before a chatbot reads your message, it chops the text into tokens, and everything it writes back is produced one token at a time. Tokens are also how AI companies measure usage, set limits and calculate prices.
What a token looks like
A token is not quite a word and not quite a letter. Common short words, such as "the" or "cat", are usually a single token. Longer or rarer words are split into pieces: "unbelievable" might become "un", "believ" and "able". Punctuation, spaces and numbers count too, and an emoji or a word in an unusual script may take several tokens.
The software that does this splitting is called a tokeniser. It is built once, in advance, by finding the pieces that appear most often in huge amounts of text. Each token is then given a number, and the model works with those numbers. If you want to know how a model uses them, our guide to how large language models work explains how it predicts the next token, over and over, to produce a reply.
How to estimate tokens
OpenAI's help centre gives these rules of thumb for English text:
- 1 token is about 4 characters.
- 1 token is about three-quarters of a word.
- 100 tokens is about 75 words.
So a 750-word essay is roughly 1,000 tokens, and a long novel runs to well over 100,000. These are estimates. Different models use different tokenisers, so the same text can come out at slightly different counts, and OpenAI notes that other languages can differ. Companies offer free tools that count tokens exactly.
Why companies charge per token
Running an AI model costs real computing power, and the work grows with the amount of text. Tokens are a neat, fair measure of that work. Developers who build apps on a model through an API (a way for software to talk to another service) normally pay per million tokens.
Some details are worth knowing, all from OpenAI's own explanation of its pricing:
- Input and output are priced separately. What you send and what the model writes back usually have different rates, and output tends to cost more.
- Reasoning tokens are billed. Some models "think" before answering. OpenAI says these hidden reasoning tokens "count toward output usage and are billed as output tokens", so a short answer can cost more than it looks.
- Cheaper per token is not always cheaper overall. Models split text differently and may write different amounts, so total cost can differ from the headline price.
We don't quote prices here because they change often. Check the provider's pricing page for current rates.
Tokens also set limits
Tokens limit how much a model can handle at once. This limit is the context window: the total number of tokens the model can look at in one go, including your question, any documents, the conversation so far and its own reply. Anthropic's documentation describes it as the model's "working memory". Our guide to the context window goes into this.
Tokens also explain some everyday behaviour:
- A long chat can hit a limit or start to "forget" early parts.
- Pasting a big document uses up much of the window.
- Usage caps on subscriptions are often tied to how many tokens you use.
What it means for you
If you only use a chatbot app, you rarely need to count tokens. But knowing about them helps in three ways.
- Shorter, clearer prompts waste less. Cut copied text you don't need.
- Start a new chat for a new topic. Old messages keep taking up space.
- Know what is billed if you use an API. Costs can rise quickly if an app sends long documents or runs many steps. Set spending limits.
Tokens are also part of why AI models can make things up: the model predicts likely next tokens rather than looking up facts.
Read on
Common questions
How many words is 1,000 tokens?
Roughly 750 words of ordinary English, going by OpenAI's rule of thumb that 100 tokens is about 75 words. It varies with the text and the model.
Do I pay per token when I use ChatGPT or Claude?
Usually not directly. Consumer apps tend to charge a monthly subscription or have free tiers with usage limits. Per-token pricing mostly applies to developers and businesses using an API, though the limits behind subscriptions are also based on usage.
Why do other languages use more tokens?
Tokenisers are usually built mostly on English text, so other languages and unusual scripts can be split into more, smaller pieces. The same sentence can therefore cost more tokens. OpenAI notes that its estimates differ by language.
Sources and further reading
Spotted a mistake or something out of date? Tell us and we will fix it. Read how we research and write our articles.
Read next
What is a context window? AI's working memory explained
A context window is how much text an AI model can consider at once. Learn what counts toward it, why long chats go wrong, and how to get better answers.
Updated 22 November 2026
How ChatGPT works: large language models explained simply
How do chatbots like ChatGPT, Claude and Gemini actually work? A plain-English explanation of large language models: tokens, training, prediction and why they sometimes get things wrong.
Updated 22 November 2026
What is artificial intelligence? A plain-English explanation
What AI actually is, how it learns, what it can and can't do, and why it suddenly seems to be everywhere. No jargon, no maths.
Updated 22 November 2026