You pay for what you send
A 128K context pack billed on every request is simply more input tokens than an 8K pack. Long windows enable the use case; they do not make those tokens free.
Fit is not the same as affordability
A model may accept 1M tokens and still be the wrong economic choice. Compare pack cost across models that fit, then decide whether RAG or caching shrinks the pack.
Reserve room for output
Context + completion must stay under the window. Always leave headroom for the answer, tool calls, and retries.
FAQ
Should I always use the largest window?
No. Prefer the smallest pack that keeps quality, then use retrieval or prompt cache for the rest.