What is a token in AI? The 4 kinds and what each one costs

What is a token?

A token is the smallest piece of text an AI model works with. It can be a whole word, part of a word, a single character or a punctuation mark.

The model never reads your words. It splits your text into tokens, works on those, and writes its answer back one token at a time.

For example, a short, common word like "the" is usually one token. A long or rare word breaks into several pieces. And code breaks into many, because every bracket and underscore counts.

Even a space changes things. In OpenAI's own example, "red", "Red" and " red" with a space in front can each be a different token.

Why does AI use tokens instead of words?

Each model has its own encoding, a fixed list of tokens it knows, each with its own ID number. Common words get a token of their own, and everything else is built from smaller pieces.

That's how a model can read any text at all: new words, names, typos, other languages and code. If it only knew whole words, every word it had never seen would be a dead end.

It's also why two models can count the same sentence differently, because each one uses its own list.

How many words is a token?

In English, one token is about 4 characters, or three quarters of a word. OpenAI and Anthropic both give that rule of thumb.

So 100 tokens is about 75 words, and 1,000 tokens is about 750 words:

It works the other way too. Multiply words by 1.33 to get tokens, so a 500-word prompt is about 670 tokens, not 500.

Files add up fast. Anthropic's own estimates put an average web page at about 2,500 tokens and a research paper PDF at about 125,000.

Three things change the count:

• Language – The 4-characters rule is for English. Other languages have their own ratio, so the same message can use more or fewer tokens.
• Code – Symbols, brackets and indentation each take tokens, so a line of code uses more tokens than a line of prose.
• The model – Each model family splits text its own way. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text than earlier Claude models.

The four kinds of tokens

When you use AI through an app or an API, your tokens fall into four kinds, and each one is counted and priced on its own.

Input tokens

Input tokens are everything you send: your message, your files, your rules and the whole conversation so far.

That last part is the one people miss. Every new message sends the full history again, so your 40th message resends the first 39 with it.

That's why a chat that has run all afternoon costs more per question than the same question asked in a fresh one.

Output tokens

Output tokens are what the model writes back, and they're the expensive half. On Claude, they cost five times as much as input tokens.

So asking for a summary instead of a full rewrite, or for just the lines that change instead of the whole file, saves more than trimming your prompt.

Cached tokens

Cached tokens are input the provider has already processed and can reuse, as long as the start of your request hasn't changed.

Your rules and context files are the same at the start of every message, so they get stored and charged at a fraction of the price. On Claude, reading from the cache costs a tenth of the normal input price or less, and the cache lasts 5 minutes, renewed each time it's used.

The catch is that the cache only matches from the start. Change one character near the top, like a date or a timestamp, and everything after it is charged in full again.

Reasoning tokens

Reasoning tokens are the thinking a reasoning model does before it answers. You don't see them, but they count, and they're billed as output tokens.

So a three-line answer can still use a lot of tokens if the model thought hard to get there.

What do tokens cost?

API prices are per million tokens, and input and output each have their own price. These are Claude's, from Anthropic's pricing page on October 1, 2026:

A single answer is cheap. A 1,000-word article is about 1,330 output tokens, which costs about 1.3 cents on Claude Sonnet 5.

The cost comes from repetition. Say an agent has 100,000 tokens of files and history in its context, and you ask it 30 more questions. That's 3 million input tokens, or $6 on Sonnet 5. With caching, most of those are read at a tenth of the price, and the same questions cost under $1.

On a flat plan like Claude Pro or Max, you don't pay per token, but tokens still drive your usage limit. Anthropic lists message length, file size, conversation length, model choice and effort level among the things that make you reach it faster.

I'm on Claude Max, and my live usage card above shows what the same tokens would have cost on the API.

How to count tokens

You don't have to guess. Here's how to see the real number:

• OpenAI's Tokenizer – Paste any text into OpenAI's Tokenizer to see how it splits and how many tokens it is.
• Claude's token counting – Anthropic's API has a token counting endpoint that counts a full request, files and tools included, before you send it.
• Claude Code – Type /context to see what's filling your context window, from your rules files to the conversation itself.
• The usage report – After every request, the API returns the exact number of input and output tokens it used.

How to use fewer tokens

Most of the savings come from a few habits:

Common token mistakes

These are the token mistakes I see most:

• You count words as tokens. A 500-word prompt is about 670 tokens.
• You keep one chat open all day. Every new message resends the whole conversation.
• You ask for a whole file back when you only need one change. Output costs five times as much as input.
• You put a date or anything else that changes at the top of a prompt. That breaks the cache for everything after it.
• You compare models only on their price per million tokens. Models split the same text differently, so compare what a real task costs.

Final thoughts on tokens

Tokens are the number AI tools run on, even when the app hides it from you.

You don't need to count every one. Remember the rule of thumb, about three quarters of a word, and remember where tokens go: long chats, big files and long answers.

Keep reading:

• What is Claude Code? – the AI agent I use every day, and what it costs
• How to build an AI OS – set up your context once, so you stop pasting it into every chat
• The AI glossary – the AI terms you'll meet most, in plain English

Next time a chat starts forgetting things or your plan runs out early, check the tokens first.