Rerank is a trade: extra calls vs less context
Without rerank, many stacks stuff all K retrieved chunks into the generator. With rerank, you score those chunks and keep N≪K. The bill goes up on the reranker and usually down on generation input.
Listwise vs pointwise
Listwise scoring sends all K chunks in one prompt (one call, large input). Pointwise scores each chunk separately (K calls, smaller inputs). Pointwise explodes quickly as K grows.
Quality is out of band
CentsPerToken prices the API path. Whether rerank improves answer quality is an eval question—pair this tool with your LLM-as-judge suite.
FAQ
What about Cohere / Voyage rerank APIs?
Those are often cheaper per query than a general LLM. Use this calculator as an LLM-rerank bound, then compare to the vendor’s per-search pricing.