If you're comparing Cohere's rerank price with Voyage's, you're comparing two different units, and the shape of your candidate set decides which one is cheaper. Cohere bills per search, Voyage bills per token. For most enterprise RAG traffic, buy Voyage rerank-3 (or rerank-3-lite if latency is tight). If your documents can't leave your network, or you run more than roughly half a million queries a month, self-host the Apache-licensed bge-reranker-v2-m3. Cohere Rerank 4 undercuts rerank-3 in one narrow case, a candidate set of close to 100 short chunks, and long chunks make it the most expensive option in the table. Before you buy any of them, measure first-stage recall, because no reranker can promote a document your retriever never returned.
All prices below were checked on the vendor or reseller page on October 15, 2026. The workload is the same for every column: a 30-token query and the candidate counts shown.
| Shape (per 1,000 queries) | Cohere Rerank 4 Fast | Cohere Rerank 4 Pro | Voyage rerank-3 | Voyage rerank-3-lite | bge-reranker-v2-m3 on Pinecone |
|---|---|---|---|---|---|
| 20 chunks × 300 tokens | $2.00 | $2.50 | $0.33 | $0.13 | $2.00 |
| 100 chunks × 400 tokens | $2.00 | $2.50 | $2.15 | $0.86 | $2.00 |
| 150 chunks × 1,000 tokens | ~$10.00 | ~$12.50 | $7.73 | $3.09 | $2.00 (no published doc or token limit) |
| Billing unit | 1 query + up to 100 docs of ≤500 tokens | same | tokens | tokens | requests |
| Context per pair | 32K | 32K | 32K | 32K | 512 in the reference code |
| Weights you can self-host | No | No | No | No | Yes, Apache 2.0 |
How Do Per-Search and Per-Token Rerank Prices Convert?
A reranker is a cross-encoder: it reads the query and each candidate document together and returns a relevance score, which you use to reorder the top-K results from your vector or keyword search before they reach the model. Both billing schemes charge for that scoring work, and they count it in different units.
Cohere defines a search unit on its pricing page as "one query with up to 100 documents to be ranked." Any document that runs past 500 tokens, counting the query, is split into chunks, and every chunk counts as a document toward the 100. A request with 101 short documents is therefore two searches, and a single 1,500-token document is three or four documents. Amazon Bedrock's pricing page describes the same rule for Cohere Rerank 3.5 and gives a worked example: 350 documents bill as 4 queries.
Voyage bills tokens. Its pricing page gives the formula: the query's token count multiplied by the number of documents, plus the tokens in all the documents. rerank-3 and rerank-3-lite cost $0.05 and $0.02 per million tokens, the same rates as rerank-2.5 and rerank-2.5-lite, and new accounts get 200 million free tokens.
The conversion falls out of the formula. One Cohere search at most holds 100 documents of about 470 tokens (500 minus a 30-token query), so a full search is roughly 50,000 Voyage tokens, or $0.0025 on rerank-3. That's the Rerank 4 Pro price for the same search. A Cohere search with fewer documents, or shorter ones, still costs a full unit, while the Voyage bill shrinks with it.
Where Does Each Pricing Model Win?
Per-search billing only wins when you pack close to 100 short chunks into every call. At 100 chunks of 400 tokens, Cohere Rerank 4 Fast at $2.00 per 1,000 searches comes in just under Voyage rerank-3 at $2.15. That's the single row in the table where Cohere is cheapest among the frontier models, and rerank-3-lite still costs less than half of it at $0.86.
The rates for each column, all checked October 15, 2026:
- Cohere's own pricing page doesn't list a per-search API rate for Rerank 4. It defines the search unit and lists dedicated Model Vault instances. Pinecone's pricing page resells cohere-rerank-4-fast at $2 per 1,000 rerank units, and Orq.ai's model page lists Rerank 4 Pro at 2.5 per 1,000 searches. Bedrock still sells Rerank 3.5 at $2.00 per 1,000 queries. Confirm the direct rate in your Cohere dashboard before you sign anything.
- Voyage's rates come from its pricing page.
- Pinecone hosts bge-reranker-v2-m3 at $2 per 1,000 requests on Standard and Enterprise, with 500 free requests a month on Starter.
At 20 short chunks, per-search billing charges for 80 empty slots. Voyage rerank-3 costs $0.33 per 1,000 queries against Cohere's $2.00 to $2.50, which makes Cohere six to eight times more expensive for a reranker that sees only 20 candidates. The "retrieve 20, keep 5" setup common in tutorials sits in this row.
At 150 chunks of 1,000 tokens, the 500-token cap does the damage. Each document with its query splits into about three chunks, so 150 documents become about 450 and bill as five searches: about $10.00 per 1,000 queries on Fast and $12.50 on Pro, against $7.73 on rerank-3 and $3.09 on rerank-3-lite. The Cohere figures in that row are estimates, because Cohere doesn't publish its exact split boundary. Your search_units field in the Rerank API response reports the real count, so log it for a day before you budget.
Scaled to one million queries a month, the 100-chunk shape costs $2,500 on Rerank 4 Pro, $2,150 on rerank-3 and $860 on rerank-3-lite. The long-chunk shape costs about $12,500, $7,725 and $3,090.
What Do the Context Limits Do to Your Chunking?
Both frontier vendors now accept 32K tokens per query-document pair, so the model limit no longer forces small chunks. Cohere's billing still does. Voyage's reranker docs list 32,000 tokens of context per pair for rerank-3, up to 8,000 of them query, 1,000 documents a request and 600K tokens a request. Vercel's model listing gives Rerank 4 Pro a 32,000-token context, and VentureBeat reported that this is four times the window of Rerank 3.5.
On Cohere the API defaults max_tokens_per_doc to 4,096 and truncates beyond it, but the pricing FAQ still chunks at 500 tokens for billing. A long-document corpus can use Rerank 4's context, then pay for it by the 500-token slice. On Voyage, a 4,000-token contract clause costs exactly its tokens and nothing more.
The open models are tighter. The bge-reranker-v2-m3 model card truncates at max_length=512 in its Transformers example, which suits 300 to 400-token chunks and nothing longer. Qwen3-Reranker ships in 0.6B, 4B and 8B sizes with a 32K context under Apache 2.0, if you need long pairs on your own hardware.
Which Open-Weight Reranker Can You Legally Run in Production?
bge-reranker-v2-m3, mxbai-rerank-v2 and Qwen3-Reranker are Apache 2.0. Jina's current rerankers aren't. Jina's reranker page lists jina-reranker-v3.5, v3, m0 and v2-base-multilingual under CC BY-NC 4.0 and says "commercial production use requires a commercial license," which Elastic sells as Jina On-Prem. The hosted Jina API and the official cloud marketplace images are covered. Weights you pull from Hugging Face onto your own GPU are not.
That makes Jina the loser for self-hosting. A team that benchmarks jina-reranker-v3.5 on a laptop, likes the numbers and ships the container has shipped an unlicensed model. If you want Jina, buy the API or the licence. Jina's public pages don't show a per-token API rate outside the key dashboard, so it isn't in the cost table.
The licensed open options:
- bge-reranker-v2-m3 is a 0.6B multilingual cross-encoder on top of bge-m3. It's the safe default for a self-hosted baseline.
- mxbai-rerank-large-v2 lists 0.89 seconds of latency on an A100 and a BEIR average of 57.49 on its own model card. The card doesn't state the candidate count behind the latency figure, so treat it as a vendor number.
- Qwen3-Reranker-8B is the choice when quality matters more than GPU cost, and Voyage's own comparison puts rerank-3 3.02% ahead of it.
What self-hosting costs and adds at p95
One NVIDIA L4 (an AWS g6.xlarge) is $0.8048 an hour on demand, per Vantage's EC2 price index, or about $587.50 for a 730-hour month. Two for redundancy cost about $1,175 a month. That pays for itself against Voyage rerank-3 at roughly 546,000 queries a month on the 100-chunk shape and about 152,000 on the long-chunk shape. On the 20-chunk shape it takes more than 3.5 million. Those crossovers assume one L4 can keep up with your traffic, so load-test before you count on them.
Published latency numbers for self-hosted rerankers disagree by an order of magnitude, because hardware, batching and candidate count all change the answer. In a September 2026 test by an independent developer, who also promotes a competing model, hosted APIs reranking a BM25 top-30 measured 287 ms p95 for Voyage rerank-3, 456 ms for Cohere rerank-v3.5 and 607 ms for Jina reranker-v3.5, network included. bge-reranker-v2-m3 on an Apple M-series GPU with no batching took 2.2 seconds and scored a lower nDCG@10 than all three hosted APIs, 0.471. A freelance AI engineer's comparison reports 80 to 150 ms for BGE on a T4 GPU with 20 documents. Measure p95 on the GPU you'll actually deploy, with your real candidate count.
Who Produced the Quality Claims?
The headline accuracy numbers in this market come from the vendors that win them. Voyage's rerank-3 announcement claims rerank-3 beats Cohere Rerank v4.0 Pro by 2.72% NDCG@10 across 95 datasets, and by 13.84% on long documents. Voyage ran every evaluation itself, reranking up to 100 candidates from four first-stage retrievers, and the post lists no independent verification.
A 2.72% average gain on public datasets says little about your corpus of insurance policies or Jira tickets. The independent test above is useful mostly for its ordering on three BEIR sets, and it reranked only 30 candidates.
So run your own. Write 200 real questions, each labelled with the chunk IDs that answer it; our guide to hand-writing a RAG eval set covers how. Pull the top 100 from your current retriever once and store them. Send the same stored candidates to each reranker, so the only variable is the reranker. Score nDCG@10 or hit@5, log p95 latency from the region you deploy in, and log what each call billed. At the 100-chunk shape the whole Voyage run is about 8.6 million tokens, well inside the 200 million free. The Cohere run is 200 searches. JPMorganChase showed that pooling LLM-judge labels across candidates keeps a bake-off like this cheap.
Will a Reranker Fix Your Retrieval at All?
Only if the right chunk is already in the candidate set. A reranker reorders the top-K your first stage returns, so recall at K is the ceiling on everything it can do. If the answer sits at position 340 and you rerank 100, no model will find it.
Anthropic's contextual retrieval write-up shows both halves. Contextual embeddings plus BM25 cut its top-20 retrieval failure rate from 5.7% to 2.9%. Adding a Cohere reranker over the top 150 brought it to 1.9%. Reranking helped, but most of the gain came from fixing what the first stage returned.
Measure recall at 20, 50, 100 and 150 on your 200 questions before you price a reranker. If recall at 100 is already high and the answers are simply ranked too low, a reranker is the right purchase. If the answer chunk is missing on more than about one question in ten, spend the money on hybrid search, better chunking or a better embedding model first. Permission filters and OCR errors also remove documents before reranking runs, and no reranker can see them.
Who Should Not Pick Each Option?
Skip Cohere Rerank 4 if your chunks run past about 470 tokens, or if you rerank fewer than 50 candidates. Either way you pay for capacity you don't use. It fits a team already on Azure or Pinecone that wants one invoice and sends close to 100 short chunks per query. Above about 1.6 million searches a month, its dedicated Model Vault instance at $3,250 a month for Rerank 4 Fast costs less than the $2.00 per 1,000 rate.
Skip Voyage rerank-3 if your security team won't approve another processor for document text, or if a vendor-run benchmark is the only reason you're switching. The rerank-3 post credits "the Voyage team at MongoDB," which matters if MongoDB is already a vendor you're consolidating or avoiding.
Skip self-hosted bge-reranker-v2-m3 if you run under a few hundred thousand queries a month and nobody on the team owns a GPU service. Two idle L4s cost more than the Voyage bill at that volume, and the model scored below every hosted API in the one independent test above. Also skip it if your chunks exceed 512 tokens.
Skip Jina's open weights unless you have bought the commercial licence.
How to Decide, and What Changes the Answer
Three numbers predict regret, and you can get all three from logs in a week:
- Candidate count and chunk length decide whether per-search or per-token billing is cheaper. Read them from your retriever logs; design docs are usually out of date.
- Recall at your candidate depth tells you whether a reranker can help at all. If recall is the problem, a reranker is wasted spend.
- Monthly query volume decides build or buy. Below about 500,000 queries a month at 100 chunks, buy an API. Above it, price two GPUs against the token bill.
The answer changes if Cohere publishes a per-token rate, if your chunks get longer (which shifts spend toward Voyage and self-hosting), or if your data-residency requirements tighten (which makes self-hosting mandatory whatever the price).
This Week: Log candidate count, average chunk tokens and recall at 20, 50, 100 and 150 for a day of real queries.
This Month: Run the 200-question bake-off on stored candidates across rerank-3, rerank-3-lite, Rerank 4 Fast and bge-reranker-v2-m3, scoring quality, p95 latency and billed units.
Before Your Next Budget Cycle: If you're above 500,000 queries a month, load-test one L4 at your real shape and price the pair against the Voyage bill.
The Bottom Line
On the 20-chunk shape, the same reranking work costs $0.13 or $2.50 per 1,000 queries depending on which billing unit you sign up for, a spread of almost 20 times. The average accuracy gaps Voyage reports on its own benchmarks are about 3%. On Anthropic's numbers, fixing the first stage cut retrieval failures nearly three times as much as adding the reranker did. Pull candidate count, chunk length and recall at 100 before you open a pricing page.
Continue Reading
- RAG Eval Dataset: 200 Hand-Written Questions Beat Synthetic Ones
- OpenAI vs Cohere vs Qwen3: Re-Embedding 10M Chunks Costs $520
- JPMorgan Ranked 62 Retrievers for $800. Pool the Judgments.
- What RAG Actually Costs: $1,308 a Month at 10M Tokens/Day
- GraphRAG vs Vector RAG: The 1,000x Index Bill Is the Small One
- Pinecone vs Weaviate vs pgvector: Stay on Postgres