What are tokens, and why does every AI company charge by them?

A token is a small piece of text, often part of a word, that a language model reads and writes one at a time. AI companies charge by the token because it is the unit their models actually process, so it tracks the work done more closely than words or requests.

How does a model split text into tokens?

Before any text reaches the model, a program called a tokenizer cuts it into pieces and swaps each piece for a number. Models do not see letters or words. OpenAI's help centre says a token can be a single character, part of a word, a whole word or punctuation, and that spaces and capital letters change how text is split.

Most tokenizers today descend from a method called byte pair encoding (BPE). It was adapted for language models in a 2015 paper by Rico Sennrich, Barry Haddow and Alexandra Birch. The idea is simple: start with single characters, find the pair that appears most often in a large body of text, merge it into one new symbol, and repeat. Common words end up as one token. Rare words get built from smaller pieces, so the model can handle a word it has never seen.

How many words is a token?

In English, a token is about three-quarters of a word, by OpenAI's estimate, but the ratio changes from model to model. The companies publish rough conversions, and they all say these are estimates:

• OpenAI: for English, 1 token is about 4 characters, or about three-quarters of a word. 100 tokens is about 75 words.
• Google: its Cloud pricing page says about 4 characters make roughly 1 text token, counting spaces.
• Anthropic: it does not give a fixed ratio. Its pricing page says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text than earlier models.

That last point matters: the same paragraph can be a different number of tokens on two models, so a cheaper price per token does not always mean a cheaper task. OpenAI's own guidance makes the same point.

What are input tokens, output tokens and the context window?

Input tokens are everything you send: your question, any documents, earlier messages in the chat, and tool definitions. Output tokens are what the model writes back. Every request has these two sides.

The context window is the most tokens a model can work with in one request, input and output together. Anthropic lists a 1 million token context window and a 128,000 token output limit for each of its four current Claude models. Anything beyond the window has to be cut, summarised or split.

Do reasoning tokens cost money?

Yes. Reasoning models "think" before they answer, that thinking is made of tokens too, and it is billed as output. OpenAI says reasoning tokens are not shown as answer text but are billed as output tokens. Anthropic's documentation reports how many of the billed output tokens were internal reasoning, and Google's price list labels its output rate "response and reasoning". A two-line answer can sit on top of thousands of hidden tokens.

How much do tokens cost?

For the main models below, input costs $2 to $4 per 1 million tokens and output costs $10 to $20. Prices below are standard rates in US dollars per 1 million tokens, read from each maker's pricing page on 16 October 2026.

Notes: OpenAI's figures are its short-context rates; it charges more for long prompts. Google's are for its global endpoint, and the current Argon price is a promotion that ends on 30 November 2026. Each maker's top tier costs more: OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 are both listed at $10 input and $50 output.

In every row, output costs five times as much as input.

Cached input is the discount for sending the same text again, such as a long system prompt or a document you ask several questions about. In the table it is 5% of the normal input price. Storing text in the cache can cost extra: Anthropic charges 1.25 times the input price for a five-minute cache write, and OpenAI lists a separate cache-write rate.

Why do some languages cost more?

Because tokenizers do not split every language equally. OpenAI notes that other languages have a different ratio of characters, words and tokens than English. A 2023 study by Aleksandar Petrov and colleagues found the same text translated into different languages could differ in token count by up to 15 times. Since billing is per token, the same message can cost more in one language than another, and it uses up more of the context window.

Why does output cost more than input?

Our take: output costs more because the model writes it one token at a time, each step a fresh pass through the model, while your input can be read in one go. That is our reading of why the gap exists; none of the three pricing pages explains it. In practice it means long answers and heavy reasoning, not long prompts, usually drive the bill.

How do you estimate what a task will cost?

Count the tokens on each side, multiply by each price, then multiply by volume.

1. Count the input. Use the maker's token counter (OpenAI and Google both offer one), or the English rule of thumb: words divided by 0.75.
2. Guess the output, then add reasoning. Check the reasoning tokens reported in a few real test runs.
3. Multiply each side by its price, and use the cached rate for text you resend.
4. Multiply by volume.

A worked example on GPT-6.1 Sol: a 10,000-word report is about 13,300 input tokens, or about $0.027. A 750-word summary is about 1,000 tokens; add, say, 3,000 reasoning tokens, and 4,000 output tokens cost $0.04. That is about 7 cents per report, or about $67 for 1,000 reports. Run a small batch first, read the usage numbers the API returns, and scale from those rather than from the estimate.