NotCheap MAKE EVERY CREDIT COUNT
Analyze Your Costs

Cohere Rerank vs Jina Reranker vs Pinecone Rerank: Real Reranking Cost per 100 Million Documents in 2026

Share this guide

If your search finds the right topic but buries the useful answer halfway down the results, a reranker may be the fix. It takes a question and the candidate passages you already retrieved, then reorders them. Before you buy one, count questions, candidates per question and tokens per candidate. A headline like "100 million documents" hides the shape of the job, and the shape is what sets the bill.

Here's a concrete starting point: 1M queries, each reranking 100 short candidates, means 100M document evaluations. On Pinecone Standard, the published BGE rate gives a $2,000 reranking subtotal for that workload. Pinecone-hosted Cohere Fast also comes to $2,000 under the short-document assumptions below, but it bills in a different unit. Neither figure is a complete search-system bill or a quality comparison.

Documentation and prices were checked 31 October 2026. This is a source-based comparison with planning examples, not a hands-on retrieval benchmark. USD amounts exclude taxes and negotiated terms.

My practical rule of thumb: pick the route that fits your document lengths first, then compare prices. A cheap rate doesn't help if the endpoint quietly truncates the passage that holds the answer.

Choose the route as well as the model

The same vendor name can show up on two different services. Keep the rate and input limits tied to the route you'll actually call.

RouteWhat to countWhat to confirm before committing
Direct Cohere Rerank 4 Fast or ProThe billed units for that exact model and agreementCurrent price, search-unit definition, long-input settings and production access
Jina hosted jina-reranker-v3.5Billed tokens and the token package you purchaseEffective token rate, usable balance, hosted input limits and cash needed for the package
Pinecone bge-reranker-v2-m3Rerank requests$2 per 1,000 requests on Standard; request size and input limits
Pinecone cohere-rerank-4-fastRerank units$2 per 1,000 units on Standard; document chunking and truncation

The Pinecone pricing table publishes those two rates separately. It also lists 500 BGE requests per month on Starter and 1,000 on Builder, with BGE the only reranking model on those plans. Standard has a $50 monthly usage minimum, applied toward usage rather than added on top. Your existing database use can already count toward that minimum.

Cohere's billing guidance describes reranking in searches. Its current pricing page also offers dedicated Model Vault instances for newer models. The practical catch: a direct Rerank 4 dollar rate and exact billing-unit definition remain unverified here. Get the terms for your route, and don't assume Pinecone's price is a pass-through quote from Cohere.

Jina's hosted service uses purchased tokens shared across its APIs. The public top-up panel requires a valid API key to show the applicable package, and we haven't verified a current package price. Use your actual purchase amount and usable token allowance to work out the effective rate. Buying hosted API access and having permission to run commercial model weights yourself are separate decisions, so check the linked service and licence terms for your deployment.

Try your own BGE search workload

Change the questions and candidates to see where another request becomes necessary. This worksheet models Pinecone-hosted bge-reranker-v2-m3 on an active Standard account for one month. It assumes each query/document pair fits the model limit; enter your largest pair, not an average. The token count is your estimate, not a measurement by this tool.

Your monthly workload and quoted rates

500,000 document evaluations; requests per query: 1; requests this month: 10,000.

Reranking usage: $20.00. Account estimate: $50.00. Extra account cost from adding reranking: $0.00.
Reranking usage subtotal
$20.00
All eligible usage before minimum
$50.00
Account estimate without reranking
$50.00
Account estimate with reranking
$50.00
Extra cost from adding reranking
$0.00

The $30 of other usage is an example to replace with your own estimate. A zero extra cost can mean the existing minimum absorbs the reranking usage; the usage subtotal still matters. If you are opening a paid account, look at the full account estimate. Zero queries does not cancel an active Standard commitment.

Requests = queries × round up(candidates ÷ 100), separately for every query. Usage = requests × rate ÷ 1,000. Account estimate = the larger of eligible usage and the monthly minimum. Splitting large candidate sets and merging results remains work for your application. This does not estimate Cohere, Jina, quality, retries, long-input chunking, taxes, credits or paid support.

Reference rates checked 31 October 2026: $2 per 1,000 BGE requests and a $50 Standard minimum. Change them to your actual quote. BGE accepts up to 100 documents per request and 1,024 tokens per query/document pair. Estimates retain fractional usage charges; invoice rounding may differ. Planning only: this does not run reranking.

What “100 million documents” means in this budget

These examples count document evaluations, including passages reranked again for a different query. They don't mean 100M unique files sitting in a database. Each query has its own candidate set, and unused capacity from unrelated queries isn't pooled into one rerank request.

Assume 200-token candidates and a short 20-token query, with no retries and no truncation. BGE accepts up to 100 documents per request. Pinecone's Cohere Fast route accepts up to 250, but its billing unit covers up to 100 short documents. Keep every query separate when rounding up.

QueriesCandidates per queryDocument evaluationsBGE requestsCohere Fast units on PineconeReranking subtotal for either route
1M2525M1M1M$2,000
1M5050M1M1M$2,000
1M100100M1M1M$2,000
1M200200M2M2M$4,000
1M250250M3M3M$6,000

For BGE, requests = queries × ceil(candidates per query ÷ 100). The 200- and 250-candidate rows split each query across requests. The Cohere route can accept those candidate sets together, while charging two or three short-document units. In both cases, the subtotal is billed count ÷ 1,000 × $2. These are short-input examples, not a rule for every document length.

That's why cutting 100 candidates to 50 doesn't halve these request-based subtotals. It may still change latency or retrieval quality, so test that trade-off instead of paying for candidates just because an endpoint can accept them. One more gotcha: asking for only ten returned results with top_n doesn't mean you submitted only ten candidates. The API parameter limits the returned results, so choose the input candidate set separately.

Long inputs: billing chunks and model context are different things

Pinecone's Cohere Fast model card defines a rerank unit as a query plus up to 100 documents. Documents longer than about 500 tokens count as multiple roughly 500-token chunks. That's a billing rule for this hosted route, not the context length of the direct Cohere model.

Cohere's direct-model guidance describes Rerank 4 processing chunks based on a 32,768-token context, with reserved tokens and the query reducing document capacity. It also documents a default max_chunks_per_doc of 1. For long inputs, set the intended coverage explicitly and verify the exact endpoint behaviour. What the direct API can process doesn't tell you how it invoices.

The Pinecone route specifications have different limits. Hosted Cohere Fast defaults max_tokens_per_doc to 1,024 and allows up to 4,096, and longer content is truncated. BGE has a 1,024-token query-and-document limit, with truncate="NONE" returning an error for oversized input and END allowing truncation.

For example, hold the workload at 1M queries and 100 candidates per query, but make each candidate 600 tokens. Assuming two billed chunks per candidate on Pinecone's Cohere route, the result is 2M units, or $4,000. BGE still fits a 600-token document plus our 20-token query and uses 1M requests, or $2,000. This illustrates billing, not equal retrieval quality. Tokenizers and exact boundaries differ, so confirm billed usage with representative inputs before treating an estimate as an invoice.

Now try a 1,500-token candidate. BGE can't process the whole query/document pair within its stated limit. Pinecone Cohere Fast's default would also drop part of the document unless you raise its setting. A lower bill that comes from discarding the answer-bearing passage isn't a real saving. Decide whether to split, shorten or choose another route before you compare totals.

Put Jina into the same workload without guessing its price

Jina describes v3.5 as listwise, meaning it considers the query and candidate list together. Its documentation gives 131,072 tokens for the combined context, with auto-truncation, and warns that hosted limits can differ from architectural capacity. Budget the combined input rather than assuming a fixed document count guarantees coverage. Confirm the current hosted limit and the actual billed token count for your request shape.

For the 1M-query, 100-candidate example, a simple input-count approximation is:

1M × (20 query tokens + 100 × 200 document tokens) = 20.02B tokens.

This assumes the query is counted once per request and excludes formatting, internal batching or other billing adjustments. It isn't a measured Jina usage result, so replace it with the returned billed count when you have one. The original 200-token document assumption alone isn't enough to establish the final bill.

Let P be your effective dollars per million billed tokens. Then this approximation costs 20,020 × P. It equals the $2,000 Pinecone subtotal at approximately $0.0999 per million tokens. That's a calculated break-even threshold, not a Jina tariff or a recommendation to choose it.

Package cash matters too. If a quoted package provides C usable tokens for B dollars, its effective rate is B × 1M ÷ C. If you must buy complete packages and have no balance, cash required is ceil(required tokens ÷ C) × B. Check expiry, existing balance and the actual purchase terms, since unused tokens aren't a cash refund. We haven't accessed your account or made a purchase.

Build the bill around a useful search result

For a Pinecone Standard planning budget, add eligible database, embedding and reranking usage, then apply the $50 minimum once. For example, an assumed $30 of other eligible monthly usage plus $10 of reranking gives a $50 total, while $30 plus $40 gives $70. These are examples of the minimum, not quotes for a database configuration. Paid support, tax and separately priced services need their own treatment.

If you already have an index, start with the questions your users actually ask and label the passages that answer them. Compare retrieval with and without reranking, using the same questions and a clear relevance measure. Record latency, failures, truncated inputs and billed usage as well. Vendor benchmark results don't establish the improvement on your collection, and this guide promises no universal percentage gain or quality winner.

In real-world production, a small labelled test set will tell you more than any price sheet. If reranking visibly lifts the answers your users care about, it's usually worth paying for, and if the gain is hard to see, keep that budget for something else.

Use the embedding budget for document and query vectors and the vector database comparison for a configured index. A managed RAG service can include a different retrieval and answer-generation path. Choose the arrangement that meets your relevance target with a bill you can reproduce.

PASS IT ON

Useful for your team?

Share this guide with a colleague.