Rerank cost calculator

Tool ·

Hosted rerankers do not bill the same way. Cohere Rerank 4 is priced per search — one query plus up to 100 documents, with any document over 500 tokens counted as several. Google and Amazon Bedrock charge per query of up to 100 documents, and Pinecone per request of up to 100, whatever their length. Cohere Rerank 5, Voyage, Jina, ZeroEntropy, Contextual AI and Alibaba Cloud charge per token — Jina counting one request's tokens, the query plus every candidate — and so do the yuan-priced services in mainland China. Which is cheapest depends on how your passages round. Put your own numbers in.

Your workload

Monthly cost

Estimated monthly rerank cost by vendor for the workload entered above
Option Billing unit Billable volume Monthly cost

Priced in yuan, for mainland China

Estimated monthly rerank cost in yuan for the workload entered above
Option Billing unit Billable volume Monthly cost

Why the ranking changes

Cohere Rerank 4 and Voyage both end up charging for text, but they round it differently. Cohere bills each candidate as one document per started 500 tokens — query included — and each search covers 100 documents. Voyage bills exactly the tokens you send, counting the query once per document.

At list prices the two meet exactly: 500 tokens costs the same on Cohere Rerank 4 Pro ($0.0025 per search of 100 documents) and Voyage rerank-3 ($0.05 per 1M tokens) — Voyage’s own pricing page uses that very example. What separates them is rounding:

Rerank 5 removes the rounding: Cohere bill it per token, at the same list prices as Voyage — $0.05 per 1M tokens for Pro against $0.05 for rerank-3, and $0.02 for Fast against $0.02 for rerank-3-lite. Cohere don’t say how the query is counted, so the calculator counts it as Voyage does, and the two pairs come out level. Choose between them on quality, not price.

Jina charges the same $0.05 per 1M tokens, but counts them differently again: v3.5 is listwise, so the query and every candidate share one 131K-token context and the query is paid for once per query, not once per document. On a long candidate list that is a small saving — and it is not why you would pick Jina. What it buys you is the listwise comparison itself; the tokens are the cheaper part of the story.

Correction, October 2026: until now this page said Cohere’s 500-token splitting was a myth and that per-search pricing ignores passage length. Cohere’s own pricing FAQ says otherwise: a document over 500 tokens, counting the query, is split, and every chunk counts toward the 100 documents in a search. The calculator now models that, which removes the long-passage advantage the old version showed for Cohere.

Your workload: —

What about self-hosting?

Self-hosting has no per-call price, so it does not belong in the table above — you trade a usage bill for a GPU bill plus operations. The comparison only becomes meaningful at your own volume: divide your monthly GPU cost by the number of queries above and compare that to the per-query figures in the table. Below a few hundred thousand queries a month an API is usually cheaper than a dedicated GPU; well above that, self-hosting a Qwen3-Reranker or bge-reranker-v2-m3 starts to pay off. Our self-hosting guide covers the serving side.

Note that a smaller model is not automatically a worse one — Jina Reranker v3.5 reaches 63.20* BEIR nDCG@10 at 0.6B in Jina's own test, level with a 4B model, so the cheapest thing to serve may also be among the most accurate.

Assumptions and sources

Cost is only half the decision

Quality is the other half. Run a cross-encoder against your own passages in the browser — no key, no spend.

Open the live demo →

Keep reading