Cost as documents-per-query grows
Same rates as above, average documents reranked per query varied across a representative range (queries/month and token sizes held fixed). Watch the "cliff" every time doc count crosses a multiple of 100 for Cohere, versus the smooth per-token climb for Voyage AI and Jina AI.
| Docs / query | Cohere | Voyage AI | Jina AI | Cheapest |
|---|
Two fundamentally different reranker billing shapes
Cohere Rerank meters usage in whole "search units": one search unit covers one query plus up to 100 documents, with each document capped at 500 tokens. If a query sends more than 100 documents, or any document exceeds the 500-token cap, the overflow rounds up into an additional whole search unit β so reranking 101 documents costs exactly the same as reranking 200. The formula this calculator uses is effectiveDocs = docs Γ ceil(tokensPerDoc / 500); units = ceil(effectiveDocs / 100); searches = queries Γ units; cost = searches / 1000 Γ rate (rate β $2.00/1,000 searches for Rerank 3.5 and Rerank 4 Fast, β$2.50/1,000 for Rerank 4 Pro). Voyage AI and Jina AI instead bill every token actually processed β query text plus every candidate document's tokens β with no bundled headroom and no rounding cliff: tokens = queries Γ (docs Γ tokensPerDoc + queryTokens); cost = tokens / 1,000,000 Γ rate (Voyage rerank-2.5 β$0.05/1M tokens, rerank-2.5-lite β$0.02/1M tokens; Jina jina-reranker-v2-base-multilingual β$0.02/1M tokens, jina-reranker-v3.5 β$0.05/1M tokens).
The practical effect: most RAG pipelines only rerank a shortlist of 10-50 candidate chunks per query, well under Cohere's 100-document ceiling β which means you're paying for a full search unit sized for up to 100 documents even though you're using a fraction of it, and the token-based vendors usually come out cheaper. Cohere becomes more competitive once a pipeline consistently reranks close to the full 100-document cap, since at that point you're actually using the headroom you're paying for. Watch document counts that sit just above a multiple of 100 (101, 201, 301β¦) β that's where Cohere's cost jumps hardest relative to the token-based vendors. Run your own query volume and document count through the calculator above instead of comparing headline per-unit or per-token rates in isolation.
How this calculator works
The Cohere vs Voyage AI vs Jina AI Reranker Cost Calculator compares two fundamentally different reranker-API billing shapes at your monthly query volume and document count: Cohere bills in whole "search units" (one query plus up to 100 documents, each capped at 500 tokens, with overflow rounding up to another full unit), while Voyage AI and Jina AI bill per token actually processed with no bundled headroom. Cohere's cost is effectiveDocs = docs Γ ceil(tokensPerDoc/500); units = ceil(effectiveDocs/100); cost = queries Γ units / 1000 Γ rate; Voyage AI and Jina AI are both cost = queries Γ (docs Γ tokensPerDoc + queryTokens) / 1,000,000 Γ rate, differing only in their per-million-token rate. The comparison table below sweeps a representative range of documents-per-query so you can see exactly where Cohere's rounding cliffs land relative to the smooth token-based curves.
These are representative pricing shapes based on typical published rates as of 2026 (Cohere Rerank 3.5 / Rerank 4 Fast β$2.00 per 1,000 searches, Rerank 4 Pro β$2.50 per 1,000 searches, each search covering up to 100 documents capped at 500 tokens each; Voyage AI rerank-2.5 β$0.05/1M tokens and rerank-2.5-lite β$0.02/1M tokens with 200M free tokens included; Jina AI jina-reranker-v2-base-multilingual β$0.02/1M tokens and jina-reranker-v3.5 β$0.05/1M tokens) β not live vendor quotes. Actual pricing varies by model version, volume commitments and negotiated enterprise terms. Always confirm current pricing directly with each vendor before budgeting.
Frequently asked questions
Is Cohere Rerank cheaper than Voyage AI or Jina AI?
It depends entirely on how many documents you rerank per query relative to Cohere's 100-document search-unit cap. Cohere bills in whole "searches" β one search covers a query plus up to 100 documents (each capped at 500 tokens) β so if you only send 20-30 documents per query, you're paying for a full search unit you're not using, and Voyage AI ($0.05/1M tokens for rerank-2.5) or Jina AI ($0.02-0.05/1M tokens) usually come out cheaper because they only bill the tokens you actually send. Once you're consistently reranking close to 100 documents per query, Cohere's $2.00-$2.50 per 1,000 searches can become competitive or cheaper again.
Why does Cohere charge per search instead of per token like Voyage AI and Jina AI?
Cohere's Rerank API meters usage in "search units": one search unit covers one query plus up to 100 documents, with each document capped at 500 tokens. If you send more than 100 documents, or any document exceeds 500 tokens, the overflow rounds up into additional whole search units β so sending 101 documents costs the same as sending 200. Voyage AI and Jina AI instead bill per token actually processed (query text plus every document's tokens), which scales smoothly with no rounding cliffs, but means cost keeps climbing linearly however many tokens you send with no bundled headroom.
What happens if I rerank more than 100 documents per query with Cohere?
Cohere rounds up to the next whole search unit. Reranking 101-200 documents in one query bills as 2 search units even though only 1 extra document pushed you over the 100-document cap; 201-300 documents bills as 3 units, and so on. The same rounding applies per-document if any single document exceeds the 500-token cap β a 1,200-token document counts as 3 documents toward the 100-document ceiling. This step-function pricing means the marginal cost of one extra document can jump from near-zero to a full search unit's price.
How much do Voyage AI and Jina AI reranker APIs cost per million tokens?
Voyage AI charges $0.05 per 1M tokens for rerank-2.5 and $0.02 per 1M tokens for the lighter rerank-2.5-lite model, with 200 million free tokens included per account. Jina AI charges $0.02 per 1M tokens for jina-reranker-v2-base-multilingual and $0.05 per 1M tokens for the newer jina-reranker-v3.5. Both vendors bill on total tokens processed (query text plus every candidate document's tokens), with no per-query or per-search unit β so cost scales linearly and predictably with token volume rather than jumping in 100-document increments like Cohere.
Which reranker API is cheapest for a typical RAG pipeline reranking 10-50 documents per query?
For most RAG retrieval pipelines that rerank a shortlist of 10-50 candidate chunks per query β well under Cohere's 100-document search-unit cap β Jina AI's jina-reranker-v2-base-multilingual ($0.02/1M tokens) or Voyage AI's rerank-2.5-lite ($0.02/1M tokens) usually cost less than Cohere, because you're only billed for the tokens you send instead of a flat per-search charge sized for up to 100 documents. Cohere becomes more competitive once your pipeline consistently reranks closer to the full 100-document ceiling per query. Run your own query volume and document count through the calculator above to see which vendor wins at your actual scale.