Context Window

A context window is the amount of material a model can handle in one context under its supported configuration. Text models commonly express this capacity in tokens. Input and generated output may share a budget, while services can impose additional separate limits. Check the rules for the actual model and interface.

A budget is not permanent memory

For an illustrative shared limit of 8,000 tokens, an input of 6,000 leaves at most 2,000 for additional generation before other applicable limits or overhead are considered. The input can include instructions, history, retrieved passages, and tool results. Counting only the latest user question misses much of the budget.

A chat application may store more history than it sends. It can omit, summarize, or retrieve earlier content. That storage is an application capability, not proof that all previous messages are present in the current model input.

Included does not mean used correctly

A model can overlook or misapply a detail even when it fits in context. The Lost in the Middle study demonstrated position-sensitive use of information in tested long-context tasks; it is evidence against assuming uniform use of all included material, not a universal ranking of every model.

Select relevant evidence and preserve decisive conditions. If a policy exception is omitted by summarization, increasing the nominal limit alone does not restore it. Evaluate the actual input and response rather than treating context size as a correctness score.

Reference: Liu et al.: Lost in the Middle.