Technology
Token
The short answer
A token is the smallest unit of text that a language model processes — typically a word fragment of a few characters. As a rule of thumb, 1,000 tokens roughly equals 700-750 German words. Tokens matter because both a model's processing capacity (context window) and its costs are measured in them.
Why tokens instead of words?
Language models break text into tokens because it's more efficient than processing whole words: common words become a single token, while rare or compound words split into multiple pieces. German compound nouns especially ("Warenwirtschaftssystem") consist of several tokens — so German text tends to require somewhat more tokens than English text of equivalent meaning.
Everything a model reads and writes counts as tokens: the instruction (prompt), any supplied documents, conversation history, and the response. API costs (price per million tokens) and processing limits (context window size) are calculated from this total.
Practical relevance for projects
In automation projects, tokens are the cost currency: a process handling long documents per operation costs more than one working with short texts. Well-designed systems optimise this — for example by sending only relevant excerpts rather than full documents to the model (RAG), or handling simple steps with smaller, cheaper models.
Token costs in project budgeting
When calculating the business case for an automation project, tokens are the variable: number of operations × average text length × model price. The good news: prices per million tokens have fallen significantly since 2023, and for most back-office processes, raw model costs sit in the low hundreds of euros per month — usually a fraction of the labour time saved.
There are still optimisation levers: smaller models for simple steps, caching recurring instructions (prompt caching), passing only relevant document excerpts instead of full texts. In well-built systems, these optimisations are standard and often cut running costs by more than half.
Input and output tokens: two separate limits
In practice, it's worth distinguishing two types of tokens. Input tokens are everything that goes into the model — instructions, documents, history. Output tokens are the generated response. Both are charged separately, and output tokens cost significantly more with most models because the model has to generate them step by step. Producing long, verbose responses therefore costs disproportionately more.
There's also a separate limit on output length, smaller than the total context window. When a very long response is requested — say a complete report in one go — it can cut off mid-way. Good systems manage this deliberately: they request concise, structured outputs, break large tasks into steps, and have results delivered in clearly bounded formats. This saves costs, prevents truncated responses, and makes results easier to process downstream — such as when a next step needs to automatically consume the output.
Practical example
An agent processes 200 incoming emails daily, each roughly 300 words. That's roughly 80,000-90,000 input tokens per day — mere cents at current model prices. Running model costs in typical back-office processes are rarely the deciding cost factor.
Frequently asked questions about Token
How many tokens does one page of A4 text contain?
Roughly 600-800 tokens, depending on text density and language. A 20-page contract comes to around 15,000 tokens — easily processed in one go by current models.
Why do AI responses cost different amounts?
Costs depend on input and output length in tokens, plus the model chosen. Output tokens typically cost more than input tokens; larger models cost more than smaller ones.
Do I need to worry about tokens as a user?
No — that's the system architecture's job. It only becomes relevant when budgeting a project with high document volumes.
Why does a very long AI response sometimes cut off?
Output has its own token limit, smaller than the total context window. When a very lengthy response is requested all at once, it can hit that limit and stop. Well-built systems avoid this by breaking large tasks into steps.
How relevant is this for your business?
In the free intro call we look at your specific process.