Context Window

What actually counts toward the context window?

More than the conversation. Anthropic’s documentation is explicit that “everything in the request counts toward the context window: the system prompt, every message in messages (including tool results, images, and documents), and your tool definitions”, and that the output the model generates counts too, including its reasoning.

Sizes have moved fast. Current Claude models carry a 1M-token context window, where Claude Sonnet 4.5 carried 200k. If the input alone exceeds the window the API returns a 400 error reading “prompt is too long”, which is at least an honest failure: the older and worse outcome is a system that silently drops the oldest part of a conversation and answers from what is left.

Why is a bigger context window not automatically better?

Because accuracy falls as the window fills, and the vendor says so. Anthropic’s own documentation names the effect: “as token count grows, accuracy and recall degrade, a phenomenon known as context rot. This makes curating what’s in context just as important as how much space is available.”

That is an unusually candid statement from a company selling the larger number, and it is the reason to distrust any argument that a bigger window makes retrieval quality irrelevant. More room to put documents in is not the same as more attention paid to each of them.

How does any of this reach a marketing page?

Through the budget your page is competing for. An assistant answering a question fills its window with several retrieved sources, its own instructions and the conversation so far, and your page is one of the things competing for the remainder. A passage that settles a question in eighty words is cheap to include; three thousand words that eventually settle it are not.

This is the mechanical reason behind advice that usually arrives as taste. Density is not a style preference in this setting: it is what makes a passage worth the space it costs, and context rot means the pages that do get included are competing with each other for attention once they are in. It also explains something that looks like unfairness.

A short page that settles the question can be included whole, while a long page contributes one extracted passage and leaves the rest of its argument outside the window, unread, doing nothing for the reader who never sees it. The window is also shared rather than yours. Several sources are competing for the same space on every answer, so the practical question is not how much of your page could fit but how much of it is worth including once four other pages have made their claim.

Related terms

• Token (LLM)A token is the unit a language model processes text in: a fragment that may be a whole word, part of one, a space or a punctuation mark.
• ChunkingChunking is the step that splits a page into passages so a retrieval system can select one without the rest.
• Knowledge CutoffA knowledge cutoff is the date after which a model’s training data contains nothing.
• GroundingGrounding is the step that ties a model’s answer to sources it retrieved, so the reply can cite them rather than assert them.