Tokens are the small pieces of text that an AI model reads and writes, usually a short word or part of a longer word. Before ChatGPT, Claude or Gemini can read your message, it is chopped into tokens. In English, one token is on average about four characters, or roughly three quarters of a word, so 100 tokens is about 75 words. Tokens decide how much an AI can read at once and, for many services, what it costs.
An everyday analogy
Think of LEGO. You can build anything, but only from bricks. Common shapes come as one big brick; unusual shapes need several small ones.
Tokens are the AI's bricks. A common word like "the" or "letter" is usually one token. A rare or long word like "pseudonymisation" is built from several. Numbers, punctuation and spaces count too.
Why AI uses tokens instead of words
A model cannot store every word in every language, plus every typo, name and new term. Instead it learns a fixed set of building blocks, often tens of thousands of them. With those it can represent any text, even words it has never seen, by combining pieces.
This also explains a few odd things:
- Other languages often need more tokens. Tokenisers are usually built mostly from English text, so the same sentence in German, Greek or Japanese can take noticeably more tokens.
- AI is clumsy with letters. Because it sees "tokens", not single letters, questions like "how many r's are in strawberry?" have tripped models up.
You can try this yourself: OpenAI offers a free online tokenizer that shows how a sentence is split.
Why tokens matter to you
1. Cost. AI providers charge for API use per token, both for what you send (input) and what the model writes (output). Output tokens usually cost more than input tokens. A long document plus a long answer costs more than a short question.
2. Length limits. Every model has a context window: the maximum number of tokens it can consider at once. That includes your messages, its replies, attached files and hidden instructions. Our page on context rot explains what happens as you approach that limit: quality often drops well before the window is full.
3. Answer length. Models also have a limit on how many tokens they can write in one reply, which is why very long outputs sometimes stop mid-sentence.
A concrete example (fictional)
"Mira", a paralegal at a mid-sized firm, uploads a 60-page contract and asks for a summary. Roughly:
- 60 pages at around 500 words a page is about 30,000 words.
- At about 0.75 words per token, that is around 40,000 tokens.
- Her question and the AI's two-page summary add a little more.
That fits easily in the context window of current large models. But if she then uploads ten more contracts in the same chat and keeps asking questions, the conversation can grow to hundreds of thousands of tokens. Answers become slower, more expensive and less precise. Starting a new chat per contract, or using a knowledge base that fetches only the relevant pages, keeps the token count low.
These figures are estimates: the exact count depends on the model's tokenizer and the language.
Tokens and what they reveal
Tokens are not a privacy feature. Splitting "Mrs Jones, diabetes type 2" into tokens does not hide anything: the model, and anyone with access to the input, still sees the full text.
What this means for confidential data
Practical tip: sending less is safer and cheaper. Paste only the section you need rather than the whole file, and remove names you do not need.
What happens to those tokens after processing depends on the provider's terms. If a provider may reuse your input, it can de-identify it, and the remaining content falls outside the GDPR. Legally it can then be pooled and sold. We trace that route on the data trail.
Where Wysor fits
In Wysor you can choose between GPT, Claude, Gemini and more for each task, with different context windows and prices; compare them on our models page. We act only as your processor and never train on your content or sell it. Every connected model provider runs with zero data retention: your tokens are processed, then deleted.
FAQ
How many words is a token? In English, one token is on average about three quarters of a word, so 1,000 tokens is roughly 750 words. Other languages often need more tokens for the same text.
What is a context window? The maximum number of tokens a model can take into account at once, including your messages, its answers, files and instructions.
Why do AI services charge per token? Because the computing work grows with the amount of text read and written. Counting tokens is a fair way to measure that.
Do images and audio use tokens too? Yes. Models that handle images or audio also convert them into tokens, though the counting works differently. See what is multimodal AI.
Further reading
Questions about using AI with sensitive data? Reach us at [email protected] or through the in-app contact form.