Retrieved context usually dominates
Eight chunks of three hundred tokens is 2,400 tokens on every single request, billed every time. In most RAG systems that is more than half the prompt, and it is the part nobody examines because it was chosen once as a default.
The comparison table exists to make the alternative concrete. Going from k=8 to k=4 halves the largest line of the bill.
Fewer chunks is often better, not just cheaper
This is the unusual case where the cheap option and the good option coincide. Irrelevant context degrades answers: a model shown four relevant passages and eight irrelevant ones performs worse than one shown four relevant passages.
Reranking a wide candidate set down to a few is the standard way to get both. Whether the reranker pays for itself is a separate calculation.