Context window
A context window is the most text an AI model can take into account in one request, counting both what you send and what it writes back. It’s measured in tokens, and a token is roughly three quarters of an English word. Anything outside the window, the model simply can’t see.
Why it matters
The window sets what an agent can hold in mind at once: a contract, a codebase, a long conversation, the results of every tool it called. Run past it and something has to go, usually the oldest messages, and the agent starts forgetting what it was doing.
Most frontier models now accept a million tokens or more. That number is less useful than it sounds. Quality drops well before the limit, often at 60 to 70% of it, and every model tested gets worse as the input grows. Information buried in the middle of a long input is the easiest to lose. A bigger window also costs more per request, since you pay for every token you send, again, on every call.
Using it well
Put the most important material at the start and the end, keep the middle for supporting detail, and send only what the task needs. For large collections of documents, retrieving the relevant parts usually beats loading everything into the window.