# What are tokens in AI, how are they counted and why do they matter? > What tokens are in AI, how many characters or words they are, a measured example of the same service request in six languages, how photos and audio are counted, and why they matter for price and the context window. Joaquín Tenés · 2026-11-14 · https://delegic.com/en/glossary/tokens If you’ve ever seen the price of an AI model, it was in tokens: so much per million input tokens and so much per million output tokens. And if you’ve read that a model “accepts a million tokens”, same thing. The token is the unit everything in a language model is measured in. Here’s what it is, how much text it is, a measured example with a service request in six languages, and why it matters to a company. ## In short - A token is a piece of text: a short word, part of a long word or a punctuation mark. - According to Google, in English a token is about 4 characters, and 100 tokens are about 60 to 80 words. - The same service request takes 39 tokens in English, 44 in Spanish and 48 in Catalan, with the same tokenizer. - Photos and audio are counted in tokens too: a small photo is 258; a second of audio, 32, in Google’s models. - Tokens decide the price, how much fits in the context window and how many requests you can make per minute. ## What is a token? A language model doesn’t read letters or words, but tokens. [Microsoft’s documentation](https://learn.microsoft.com/en-us/dotnet/ai/conceptual/understanding-tokens) describes them as words, character sets or combinations of words and punctuation, into which a tokenizer splits the text before it reaches the model. The set of all the tokens a model knows is its vocabulary, and each one has a number. Google’s guide sums it up like this: “Understand and count tokens” guide on [Google AI Studio](https://aistudio.google.com/docs/tokens), accessed on November 14, 2026. Common words are usually one token, and long or rare ones are split into several. How a specific sentence is split, token by token and with each one’s number, is on the [LLM](/en/glossary/llm) page: there “won’t” comes out in two pieces, “won” and “’t”. ## Does it take the same in every language? No, and it shows in the price. Google’s 4-character rule talks about English words, and in other languages the count changes. We measured it: we translated a typical service request into six languages and counted its tokens with the same public tokenizer. “Good morning, the building’s boiler hasn’t fired since last night…”, in six languages 1. English152 characters39 tokens 2. Portuguese136 characters43 tokens 3. Spanish154 characters44 tokens 4. French171 characters47 tokens 5. German184 characters47 tokens 6. Catalan147 characters48 tokens In English it’s 39 tokens; in Spanish, 44, 13% more; and in Catalan, 48, 23% more, even though the Catalan text has fewer letters than the English one. English comes out at about 3.9 characters per token, Spanish at 3.5. For a single request it makes no difference; across thousands of conversations a month, 13% more tokens is 13% more on the bill for the same conversation. Not everything splits the same way. “Boiler” is a single token; the Spanish “caldera” is two, “cal” and “dera”; and the German “Heizkessel”, three. Error codes and numbers get split too: “F28” is two tokens, “F” and “28”, in all six versions. ## What about photos or calls? Models that understand more than text, [multimodal](/en/glossary#multimodal) ones, count everything in tokens too. Google’s guide spells it out for its models: an image of 384 pixels or less on each side is 258 tokens, and a larger one is split into 768 by 768 pixel tiles, at 258 tokens each. Audio runs at 32 tokens per second, so a three-minute call is 5,760 tokens, and video at about 100 per second at the default resolution. In field service that comes up every day: the photo of a nameplate a customer sends on WhatsApp, or the recording of a call you want summarized. Each has a cost in tokens, just like text. ## Why do they matter? For three reasons, which Microsoft explains on the same page. The first is **price**: generative AI services often charge by tokens, and the cost of each request depends on the input and output tokens, which can be priced differently. The second is the **limit**: each model has a maximum number of tokens it can handle at once, the [context window](/en/glossary#context-window), which adds up what it reads and what it writes. The third is **rate**: services cap tokens per minute, so a spike in conversations can hit that ceiling. ## What should a company look at? If you hire a ready-made agent, you usually won’t see tokens: you’ll see a price per conversation, per request or per month, and tokens are the provider’s business. Even so, two things are worth knowing. Usage-based pricing goes up with the size of the instructions and documents passed in with each answer, and photos and calls count for much more than a text message. If you build it yourselves on top of a model, the sum is simple: how many conversations a month, how many messages each, how many tokens the model reads for each answer, and the price per million the provider publishes. And if your customers write in a language other than English, expect somewhat more tokens than in English examples. ## Frequently asked questions ### How many words are 1,000 tokens? In English, by Google’s rule, between 600 and 800 words. In our request, 27 English words took 39 tokens, about 1.4 tokens per word, which gives roughly 700 words per 1,000 tokens. It’s an estimate from a single sentence; every text gives a different figure. ### Is a token a word? Sometimes. Short, common words are usually one token; long or rare ones, numbers and codes are split into several. A punctuation mark can also be a token. ### Is it the same as a security token? No. In AI, a token is a piece of text. The same word is used for unrelated things: the code that identifies a user in an app, or a cryptocurrency, for example. ### Why charge by tokens and not by words? Because that’s what the model actually processes: the work of computing each answer grows with the tokens it reads and writes, not with words or messages. ## Related terms - [LLM](/en/glossary/llm): the model that reads and writes in tokens. - [Context window](/en/glossary#context-window): how many tokens it can handle at once. - [Prompt](/en/glossary/prompt): the instructions read, and paid for, with every answer. - [Embeddings](/en/glossary/embeddings): the numbers that represent the meaning of a text. - [Multimodal](/en/glossary#multimodal): a model that also reads photos and audio.