Token (AI).
A token is the smallest unit of text an AI language model works with. The model splits every text into tokens: whole words, parts of words, punctuation or spaces. Tokens determine how much text a model can handle at once, how quickly it responds and what a request through an API costs.
Token (AI) — Explained in Detail
Language models do not read letters or whole sentences; they read tokens. A tokenizer splits text into pieces according to a fixed vocabulary. Common words such as 'and' are a single token, while rare or long words are broken into several parts. The large language model then computes with these tokens and generates its answer token by token as well. Every vendor uses its own tokenizer, so the same text yields a different number of tokens on different models.
As a rule of thumb, OpenAI gives roughly four characters per token for English text (OpenAI Help Center, 'What are tokens and how to count them?', retrieved 11.11.2026). German usually needs more tokens for the same content, because compound words such as 'Liegenschaftsverwaltung' fall apart into several pieces and vocabularies lean towards English. Swiss German, technical terms, columns of figures and code push the count higher still. Exact numbers come only from the tokenizer of the model in question.
Tokens limit what a model can see at once. The context window is measured in tokens and covers everything: the system prompt, the conversation so far, attached documents and the answer. Once it is full, older parts drop out or the request is rejected. With so-called reasoning models, internal thinking steps consume tokens too, even though they are not displayed.
API providers bill by token, separately for input and output, and output tokens are usually dearer. Current prices are on the vendors' pricing pages and change frequently. For cost estimates, volume therefore matters more than price per token: how long is the system prompt, how many documents go along with each request, how long are typical answers? An AI token should not be confused with a design token in web design, which stores the colours and spacing of a design system.
Keeping tokens and costs under control
The biggest levers are volume, not the model. A lean system prompt, sent with every request, saves on every answer. Instead of attaching whole manuals, a system using retrieval-augmented generation fetches only the relevant passages. Long conversations can be summarised rather than resent in full each time. And for simple tasks a smaller, cheaper model is often enough.
Which models and plans suit your tasks, and how costs are likely to develop, we work out in AI consulting. In our AI workshop we show teams how to use context and prompts economically. For a chatbot, a cost estimate based on expected conversations is part of the planning.
Example: an insurance broker estimates the cost of an internal assistant
A constructed example: an insurance broker in Lucerne is planning an internal assistant to answer questions about policy terms. The first draft sends the full terms of every product with each question, dozens of pages. The context window is nearly full, answers are slow and the estimated cost per question is high.
In the second draft the terms are split into sections and stored in a searchable form. Each question sends only the matching sections to the model, plus a short system prompt. Tokens per request drop markedly, and answers become faster and more accurate because the model is no longer searching irrelevant text. The broker measures the actual figures during a trial using the provider's counting tools rather than estimating them.
Related Page
AI consulting SwitzerlandFrequently Asked Questions About Token (AI)
It depends on language and model. For English text OpenAI gives a rule of thumb of roughly four characters per token. German usually needs more tokens for the same content, because long compound words are split. Exact figures come only from the specific model's tokenizer.
The model has to compute each output token individually, whereas input is processed in one go. Providers reflect this computing effort in their prices. Current prices are on the vendors' pricing pages.
An AI token is a unit of text that language models compute with. A design token is a named design value in a design system, such as a colour or a spacing. The two terms are unrelated.
More Terms Starting with "T"
TTFB (Time to First Byte)
TTFB measures the time from sending an HTTP request to receiving the first byte of data from the ser...
Tailwind CSS
Tailwind CSS is a utility-first CSS framework that allows developers to create designs directly in H...
Title Tag (Page Title)
The title tag is a web page's HTML page title. Google shows it as the clickable blue headline in the...
Ready for Your Project?
Apply this knowledge to your website — DLM Digital will help you.