Capacity is not usage
A context window caps the material the model can consider in one request. Selecting a model with a 200K window does not mean every request bills 200K tokens. A 2K-token prompt is billed based on its actual input usage, even if the model could accept much more.
History makes later turns larger
In a chat app, prior messages may be resent on every turn. A conversation that begins with a small prompt can become expensive because the input grows. Log input tokens by turn and set policies for summarizing, trimming or retrieving only relevant history.
Long input can change the rate
Some provider price sheets distinguish requests above a context threshold or specify different rates for particular model modes. Before estimating a 100K-token document, inspect the exact model's current pricing notes, not just the headline input rate. Output, cache reads and non-text inputs may also have separate prices.
A 100K-token document example
At an illustrative $3 per million input tokens, 100,000 input tokens would cost $0.30 before output. If the model returned 2,000 tokens at $15 per million, output would add $0.03. Repeating the full document across 1,000 requests would cost about $330 at those sample rates. Retrieval or caching may change that economics.
Design around the task
Measure answer quality with a short relevant excerpt versus the full document. If retrieval meets your quality target, it reduces billable input. If the task requires global reasoning across the entire file, a long context model may be justified. Compare model prices using your actual token mix.