What Is a Token in AI?

A token is a small piece of text that an AI language model reads and generates. A token can be a whole word, part of a word, a single character or a punctuation mark. AI models break everything you type into tokens, and tokens are how their limits and costs are measured: how much text a model can handle at once, how long its answers can be and, for businesses using AI services, how much each request costs.

What a token is

Language models do not read text the way people do. Before a model processes your message, a component called a tokenizer splits the text into tokens and converts each one into a number the model can work with. OpenAI's documentation for its open-source tokenizer puts it simply: language models "see a sequence of numbers (known as tokens)."

OpenAI describes a token as something that "can represent a character, part of a word, a whole word, or punctuation." Common words are often a single token. Longer or less common words are usually split into pieces. OpenAI's tokenizer documentation gives the example of "encoding," which is often split into "encod" and "ing," because "ing" is a common piece of English.

The model then works entirely in tokens. It reads your prompt as tokens, predicts the next token, adds it and repeats until its reply is complete. When the model is done, the tokens are converted back into readable text. The LATHIC Glossary has the short definition of a token.

Why AI uses tokens instead of words

Splitting text into word pieces solves a practical problem. There are far too many possible words, names, typos, technical terms and new slang terms for a model to store each one separately. By breaking text into reusable pieces, a model can handle words it has rarely or never seen.

Many tokenizers use a method called byte-pair encoding, which builds a vocabulary of frequently occurring pieces of text. Researchers Rico Sennrich, Barry Haddow and Alexandra Birch popularized this subword approach for neural language processing in a 2016 paper, and OpenAI describes its tiktoken library as a byte-pair encoding tokenizer. Different AI companies use different tokenizers, which is one reason token counts vary between models. The process of splitting text is called tokenization.

Tokens vs words

Tokens and words are not the same, and there is no universal ratio between them. How many tokens a piece of text uses depends on the language, the kind of content and the specific model's tokenizer.

AI companies publish rough rules of thumb for English text:

These are estimates, and even a single company's figures can shift. Anthropic says the newer tokenizer it introduced with its Claude Opus 4.7 model produces about 30% more tokens for the same text than its earlier models. By Anthropic's estimate, a million tokens holds roughly 555,000 words on its current tokenizer, compared with about 750,000 words on earlier models.

Other factors change the count too:

• Language. A 2023 study found that the same text translated into different languages can require very different numbers of tokens, up to 15 times as many in some cases. Some languages need many more tokens than others for the same content.
• Unusual words and names. Rare words, technical terms and names are often split into several tokens.
• Numbers, code and formatting. Long numbers, code, symbols and tables can use more tokens than ordinary prose.

Why tokens matter

Most people never need to count tokens, but tokens explain several things people notice when using AI.

Why chatbots have limits

Every model has a context window: the maximum amount of text, measured in tokens, it can consider at once. That includes your messages, any documents you upload and the model's own replies.

Context windows vary widely. As of September 2026, Anthropic lists a 200,000-token context window for its Claude Haiku 4.5 model and 1 million tokens for its larger current models, while Google lists an input limit of just over 1 million tokens for Gemini 2.5 Pro. Consumer apps may apply their own, lower limits.

When a conversation or document exceeds the context window, the model cannot take everything into account at once. Depending on the product, you may see an error, the app may drop or condense earlier parts of the conversation, or the model may lose track of details from much earlier.

Why answers have a maximum length

Models also have a limit on how many tokens they can generate in a single response. If a long answer stops mid-sentence, it may have reached that limit. The LATHIC Glossary covers this setting under max tokens.

Why AI services charge by the token

Businesses and developers who use AI models through an API, a way for software to connect to an AI service, usually pay per token. Prices are typically quoted per million tokens, and OpenAI notes that "input, cached input, and output tokens can have different prices." Output tokens, the ones the model generates, usually cost more than input tokens.

Prices vary by provider, model, type of token and time. For illustration, on September 10, 2026, Anthropic listed Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, and Google listed Gemini 2.5 Pro at $1.25 per million input tokens and $10 per million output tokens for prompts up to 200,000 tokens. Prices like these change frequently, and discounts or higher rates can apply in different situations.

A lower price per token does not always mean a cheaper result. OpenAI cautions that models "can tokenize the same text differently and generate different amounts of output or reasoning."

Most people using consumer chat apps do not pay per token. Services such as ChatGPT and Claude offer free and paid subscription plans with usage limits. Anthropic, for example, says every Claude plan has usage limits that reset on a rolling five-hour window.

What are tokens in ChatGPT?

Tokens in ChatGPT are the same units of text described above. The OpenAI models behind ChatGPT break your messages, uploaded content and their own replies into tokens. Tokens determine how much of a conversation or document the model can consider at once. ChatGPT subscriptions are generally not billed per token; developers who use OpenAI's models through its API are.

AI tokens vs crypto tokens

An AI token and a cryptocurrency token share a name but are completely different things.

European Union law, for example, defines a crypto-asset as "a digital representation of a value or of a right that is able to be transferred and stored electronically using distributed ledger technology or similar technology." AI tokens are nothing of the kind. The phrase "AI tokens" is also sometimes used for cryptocurrency projects with an AI theme, which are unrelated to the text units language models use.

How to estimate tokens

For everyday purposes, a rough estimate is enough. For precise counts, use the token counting tools that AI providers offer for their own models. Because tokenizers differ, Anthropic advises recounting prompts against the specific model you plan to use rather than reusing counts from earlier models.

To see how tokens fit into the way language models generate text, see What Is a Large Language Model?. For the wider category of AI that generates content, see What Is Generative AI?.