Context Window.

The unit of measurement is not characters or words but tokens. A token is a text fragment the model treats as its smallest unit: sometimes a short whole word, sometimes a syllable, sometimes a punctuation mark. As a rough rule, a thousand tokens correspond to about seven hundred words of English. German runs lower because of long compounds and umlauts, so a German text needs more tokens than an English translation of the same length. The consequence matters more than the arithmetic: everything in a request competes for the same space, and the answer needs room inside it too.

The technical sizes are no longer the bottleneck in 2026. Widely used models sit between roughly two hundred thousand and one million tokens, and some providers quote higher figures still. The more important finding comes from independent testing: the reliably usable length is considerably shorter than the advertised one. The fuller the window, the more often details are overlooked, particularly those sitting in the middle of a long input. In practice a large window is no substitute for selection. Relevant material in the right place beats a great deal of material in arbitrary order.

Then there is the cost side, which is easy to overlook. Billing is per token in both directions, and in a conversation the entire history is re-sent with every turn. A chat that keeps a sixty-page attachment in context pays for that attachment again with every follow-up question, and response times grow with it. The countermeasures are interim summaries of the conversation, caching of unchanging instructions where the provider supports it, and above all resisting the urge to include everything you happen to have.

A worked example makes the rule concrete. A company wants an assistant that answers questions about a four-hundred-page manual. The naive approach is to put the whole manual into every request. It works technically, but costs a multiple per question and produces worse answers, because the relevant passage drowns in the mass. The usual approach is retrieval: the manual is split into sections, the five to ten matching sections are selected per question, and only those go into the context. The window stays free for reasoning rather than ballast.