What is a context window?

A context window is, in Anthropic’s words, “all the text a language model can reference when generating a response, including the response itself.”^[1] The same documentation describes it as a “working memory” for the model, separate from the much larger body of data the model was trained on.^[1] Whatever you want the model to take into account on a given turn has to fit inside that window.

The window is measured in tokens, not words. A token is a chunk of text the model processes; OpenAI describes tokens as “commonly occurring sequences of characters” and offers a rough rule of thumb that “1 token is approximately 4 characters or 0.75 words for English text.”^[2] So a few hundred words of plain English is a few hundred tokens, and the page or two of text in a short note is roughly the same order. Counting in tokens matters because the limit is a token limit.

A rough mental model helps make that distinction concrete. A 1,000-word email comes out to somewhere around 1,300 tokens at OpenAI’s stated rate of roughly 0.75 words per token, and a 200,000-token window can hold on the order of 150,000 words, well past the length of a typical novel.^[2] That much headroom sounds generous, but it fills up faster than it looks, because the window has to hold more than just what you type. What else counts against the same budget is covered on the next page.