The basics · Layer 4
Tokens and context window
A model reads and writes in tokens, fragments of words; its context window sets how many it can handle at once.
A way to picture itThe context window is the model’s desk: everything needed for the answer (instructions, documents, conversation history) has to fit on it. The more cluttered it gets, the more likely the model is to miss the one sheet that matters.
A legal team loads a framework agreement and its twelve amendments into an assistant in one go. Everything fits in the window, but the answer misses a clause changed by the seventh amendment, buried in the middle. The team changes approach: it queries the model one amendment at a time, then has it consolidate the results. The answers are shorter, cheaper and easier to check.
Everything placed in the context window of a hosted model is sent to its provider: that is data processing, which needs a contract as soon as personal data are included (EU · GDPR, art. 28). The data minimisation principle applies: include only what the task requires (EU · GDPR, art. 5). The contexts you send may be kept in logs, on your side or the provider’s; how long they are kept and how they are protected fall under security of processing (EU · GDPR, art. 32).
“What context window does your tool actually use, what happens when our documents exceed it, and how long do you keep what we send?”
Because the model does not see words but fragments produced by automatic splitting. Tokens are also the billing unit for most providers, for both input and output.
About 500 pages for 200,000 tokens, according to an Anthropic estimate (United States · Anthropic, September 2024). The figure varies with language, layout and tables.
Not necessarily. Selecting the relevant passages, for instance with RAG, often gives a more accurate, faster and cheaper answer than loading everything.
Educational explainer, not legal advice. Up to date as of 29 October 2026.