What it means in practice
When you read text, you see letters and words. When an AI reads it (an LLM), it chops the text into smaller pieces called tokens.
In English the rough rule is that 1 token is about 3/4 of a word. A short word like “dog” is one token. A long word like “incomprehensible” may be split into three or four. In other languages it is worse: with heavy inflection and accented characters, models often spend more tokens per word than they do in English.
Why companies need to know this
For an employee paying 20 euros a month for ChatGPT Plus, tokens do not matter. For a company wiring AI into its processes, tokens are a synonym for money.
- What AI costs. When you connect your company system to OpenAI or Anthropic models through an API, you do not pay a monthly flat fee. You pay for consumption: for every token you send in, and every token the AI generates back.
- The memory limit (context window). Every model has a limit on how many tokens it can hold at once. A limit of 128,000 tokens means you can put roughly a 300-page book into a single prompt. Go beyond it and the model starts forgetting what came at the beginning.
- Non-English costs more. Because models split text in heavily inflected languages into more tokens than English, processing a document in one of those can cost 30 to 50 percent more than the same document in English.
An example from practice
Imagine you want AI to summarise every incoming email in your customer system. An average email is 500 tokens. At a thousand emails a day, that is 500,000 tokens daily.
Pick the most expensive model for the job and it may cost hundreds of euros a month. Pick a smaller, cheaper one, which is more than enough for a simple summary, and the bill can be a few euros.
In our company AI workshops we show not only how to use AI, but how to estimate the cost and pick the model so the automation makes economic sense.