A model does not read characters or words; it reads tokens, pieces of text about three-quarters of a word long in English and shorter in code or other scripts. The context window is the budget of tokens one request may use, and everything counts against it: the system prompt, every earlier turn of the conversation, the documents pasted in, tool results and the answer the model has yet to write. A 128,000-token window holds roughly a 300-page book; a 1,000,000-token window several.
The window is a hard limit, not a quality dial. A request that exceeds it is rejected or silently truncated depending on the client, and truncation is the worse failure because the model answers from part of the input without saying so. The reference catalogue lists each model's window and its separate maximum output, which is often much smaller than the input limit and matters for long-form generation.
Bigger is not free and not always better. Every token read is billed at the input rate, so a long window used in full costs a long window's price on every turn; cached-input pricing, where a vendor offers it, reduces the cost of re-reading the same prefix. Models also lose precision deep into a long context - the 'lost in the middle' effect - so stuffing a window is not a substitute for retrieving the relevant passages first.
A router treats the window as a constraint before anything else. The request's size is one of its signals; a target whose window cannot hold the request is not admissible, so nothing is truncated by a bad choice, and among the admissible models the usual scoring on quality and price applies - which means a short question never pays for a long-context model just because one exists. Retrieval narrowing, when configured, cuts the input to the passages the request needs before the window is even consulted.
