What takes up space in a context window?

Several kinds of text share the same window. There is the system prompt, the standing instructions that set the model’s behavior. There is the conversation history, every earlier user message and model reply. There are any documents you add, such as a pasted report or an attached file. And there is the model’s own response, which counts too. Anthropic’s documentation states plainly that the window includes “the response itself.”^[1]

Usage grows turn by turn. The same documentation describes “progressive token accumulation,” where “as the conversation advances through turns, each user message and assistant response accumulates within the context window,” and previous turns “are preserved completely.”^[1] A long chat or a big pasted document can fill the window even when no single message is large. Each turn’s input is all the prior history plus the new message, and that turn’s output then becomes part of the next turn’s input.^[1]

How much fits depends on the model, and the split is no longer a minor footnote. As of July 2026, Anthropic gives a 1M-token window by default, no special request needed, to Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, Sonnet 4.6, Fable 5, and Mythos 5; older or smaller models, including Sonnet 4.5, stay capped at 200K.^[1] A single request against a 1M-token model can also carry more attached material: up to 600 images or PDF pages, against 100 for a 200K model.^[1] Which tier a model sits in is not cosmetic — a document that fits whole in one window has to be chunked or summarized for the other.

Growth in the window has a cost beyond storage. Because attention compares every token to every other token (the why limits exist page covers the arithmetic), a new turn is not processed in isolation. Turn 50 of a long conversation attends over everything accumulated across turns 1 through 49 as well as its own new content, so both the compute cost and the latency of a reply grow as the conversation lengthens, not just the token count sitting in memory. A short exchange and a long-running one are not the same job for the model, even when the newest message is the same length in both.