Last updated: 2026-06-22

The context window explained

The context window is the maximum number of tokens a language model can process in a single call, counting both the input it reads and the output it generates. A request fits when the input plus the tokens reserved for the reply stays within this limit. Because the window is fixed for a given model, planning a call means budgeting its components against that ceiling.

The budget

The input of a single call is the sum of the system prompt, the prior conversation history that is re-sent, and the current user message. Reserving room for the reply gives the total the window must hold:

where $W$ is the context window and $n_{out}$ the reserved output. Utilization is the share of the window the request consumes:

A negative $n_{rem}$ means the request overflows the window and cannot run as composed.

Worked example

Consider a request with an 800-token system prompt, 6000 tokens of history, and a 1200-token user message, reserving 2000 tokens for the reply, against a 128,000-token window:

The request needs 10,000 tokens and leaves 118,000 free, using about 7.8% of the window. There is ample headroom: history could grow many times over before the window becomes the binding constraint.

Overflow behavior

When the total exceeds the window, providers do not silently truncate in a way the caller controls; the call is rejected or the oldest context is dropped, either of which degrades the result. The reserved output deserves particular care: if the input nearly fills the window, the model has little room left to respond and the reply may be cut off mid-generation. Reserving output explicitly prevents an input that fits from producing a truncated answer.

Application

The window is a hard limit, distinct from cost, which scales with tokens used. History is usually the largest and most compressible component, so fitting a long conversation typically means summarizing or trimming it, shortening the system prompt, lowering the reserved output, or moving to a model with a larger window. To size the input from a word-count target, convert with Tokens to Words Calculator; to see how the re-sent history drives both the window and the bill upward over many turns, see Context Cost Growth Calculator.