Why it matters
If a model has to summarize a 300-page contract, the input, instructions, and requested output all compete for room. Exceed the window and the request may fail, lose context, or require a different document strategy.
What counts toward the limit
- System and developer instructions.
- The user’s message and conversation history.
- Retrieved passages, files, or tool results.
- The model’s generated response.
- Sometimes hidden formatting or tool-call context.
Why bigger is not always better
A large window lets you send more material, but irrelevant material can bury the important passage and increase cost or latency. Good retrieval, summarization, and document selection often beat simply sending everything.