What Are AI Tokens, and Why Do They Decide What You Pay?

When you send a message to an AI model, it doesn't read your text word by word. It breaks the text into small pieces called tokens, works with those, and writes its reply as tokens too. Tokens are the currency of AI: they decide how much the model can keep in mind at once, how fast it replies and what you pay.

You don't need a technical background to understand them. Knowing the basics helps you read pricing pages, avoid surprise bills and compare tools sensibly.

What a token actually is

A token is a small chunk of text. Sometimes it's a whole word, sometimes part of a word, sometimes a single character. The text is cut up by a tokenizer, a tool that splits writing into the pieces the model learned to work with.

For common English words, a token often matches a word: "chat" is one token, "model" is one token. Longer or less common words get split into pieces, so a word like "tokenization" may become two or three tokens. Punctuation, spaces and line breaks are tokens too. Numbers are often split differently from words.

Each AI company uses its own tokenizer. OpenAI's is called tiktoken, and Anthropic and Google have their own. So the same sentence can come to a slightly different number of tokens in different tools, though for everyday English the difference is small.

As a rough guide, one token is about four letters, or three-quarters of an English word. So 100 tokens is about 75 words. OpenAI's help centre gives a similar example: 1,500 words is about 2,048 tokens (source: OpenAI help centre (opens in a new tab)).

Non-English text often uses more tokens per word, because tokenizers are usually trained mostly on English. The gap is small for languages like French or Spanish, and much bigger for languages written in other alphabets, so using them costs more. Code also uses more tokens than you might expect, because every bracket, symbol and indent counts.

A worked example

Take this sentence: "The model read the document carefully."

That's six words. A typical tokenizer turns it into about seven tokens: each common word is one token, and the full stop is one more. A rarer word would be split into two or more.

Now imagine that sentence is part of a 2,000-word document. At three-quarters of a word per token, that's about 2,700 tokens to send, before the model has written a word. If the reply is 500 words, add about 670 tokens. That one exchange comes to roughly 3,400 tokens.

AI companies price their APIs per million tokens, so at typical rates one exchange like this costs a fraction of a cent (check the provider's pricing page for the real figure). The cost starts to matter when you send thousands of requests or work with very long documents.

Why tokens determine what you pay

AI companies charge for tokens in both directions: input tokens (what you send) and output tokens (what the model writes back). Output tokens usually cost more, because writing text takes the model more work than reading it.

This affects how you use AI tools:

• If you send the same long set of instructions with every request, you pay for them every time.
• Asking for a long answer when a short one would do costs you more output tokens.
• Many APIs offer prompt caching: they remember text you send repeatedly and charge less for it the next time.

Behind the per-token price is the inference cost: what it costs the provider to run the model each time it answers. Bigger models need more computing power per answer, so they usually cost more per token.

What tokens determine beyond price

Tokens matter for more than your bill. Every model has a context window: the most tokens it can keep in mind at once, counting both the conversation so far and the reply it's writing. A 128,000-token window holds roughly 95,000 words. An 8,000-token window holds about 6,000.

When a conversation outgrows the window, the model either stops or starts losing the earliest parts. That's why very long chats can seem to "forget" what you said at the start.

The size of the window also decides what you can do:

• Summarising a 50-page contract requires enough context to hold it all.
• Writing code across multiple files works better with a larger window.
• Simple Q&A rarely comes close to any model's limit.

What tokens are often confused with

Words: Tokens aren't words. Common English words are usually one token each, but technical terms, long words and other languages use more.

Characters: A token isn't one letter either. An English letter is about a quarter of a token.

Credits or message limits: Some subscriptions hide tokens completely and count "messages" or "credits" instead. Then you can't compare them directly with an API's per-token price unless you know how they convert.

Model routing: Some services automatically send each question to a cheaper small model or a pricier large one, depending on how hard it looks. So two messages of the same length can cost different amounts.

How tokens affect API rate limits

If you use an AI model through its API, you'll meet API rate limits: caps on how much you can send in a minute. They're usually set in tokens per minute (TPM) and requests per minute (RPM). A limit of 100,000 TPM means all the tokens you send and receive in one minute can't add up to more than 100,000.

Long requests use up that limit faster, so processing a batch of long documents can suddenly slow down or fail. Most providers show your usage in a dashboard, and higher plans usually come with higher limits.

What to check when comparing AI tool costs

• Input and output prices. They're two different numbers. A model with a low input price may charge a lot more for output. Check both before deciding a tool is cheap.
• Context window size. A bigger window lets you send longer documents and keep longer conversations, but each request costs more if you fill it. A 128,000-token window is a ceiling, not what every request uses.
• Prompt caching. Some providers charge less for text you send over and over, such as the same instructions on every request. OpenAI, Anthropic and Google all offer this on at least some models; their documentation explains how to turn it on.
• Free plan limits. Apps like ChatGPT and Claude describe limits in messages or usage caps, not tokens. A long message or a file upload uses more than a short one, so you'll hit the limit sooner if you send those often.
• Token counters. OpenAI's Tokenizer (opens in a new tab) shows how many tokens any text uses, and Anthropic's API can count tokens for you. Check before you commit to a plan for heavy use.
• Price differences between models. On the same platform, the top model (often called "Pro" or "Opus") usually costs several times more per token than a lighter one (often called "Flash" or "Mini"). Many jobs run fine on the lighter model, so test it before defaulting to the biggest.

If you're comparing AI APIs on cost, start with the best AI APIs for developers, then check each provider's own pricing page for current figures.