A larger context window is not a better one. Models get less reliable as the input grows, and every token you put in the window is billed again on the next turn.
Retrieval beats stuffing
Chroma tested 18 models and found performance degrades unevenly as input length increases — and the degradation starts before the window is full. The same question answered from a short, relevant passage beats the same question answered from a long, mostly irrelevant one.
So the goal is not to fit more in. It is to put less in and keep what matters.
Price is per turn, not per conversation
A conversation of twenty turns re-sends the history twenty times. At Opus 5.5's $4 per million input tokens, a 100,000-token history costs $0.40 per turn by turn twenty, and $8 across the conversation before any output.
Two levers cut that. Keep the history short. Put the stable part — the system prompt, the reference document — at the front and cache it, because Anthropic prices a cache read at a tenth of a fresh input token.
Put the instruction last
Models attend more reliably to the end of a long input than the middle. If a long document and an instruction share a prompt, the instruction goes after the document, not before it.
Summarise on a budget
Compaction is not free and not lossless. Summarise the parts that are settled, keep the decisions verbatim, and write the file paths and error messages into a notes file rather than trusting the transcript.
Measure the window you actually use
Log input tokens per request for a week. Most teams find a long tail: a handful of requests carrying 80 per cent of the tokens, usually a whole file pasted in where a search would have done.
That tail is the cheapest thing to fix.