Cohere Rerank Pricing: Search Units Are Not Dedicated Model-Vault Hours

What counts as one search unit

Cohere’s official pricing explanation defines a Rerank search unit around one query and up to 100 documents. The documentation also explains that long candidate documents can be split into chunks, which count toward the documents handled. A naïve system can therefore underestimate reranking volume by counting only user queries while ignoring the number and size of candidate passages.

Do not multiply a text-generation cost per million tokens by the number of Rerank requests. Rerank is a separate product with separate usage semantics. For a search application, record query count, ranked documents, chunking and model version as its own metering data.

Dedicated Model Vault is capacity billing

Cohere’s Standard Vault documentation publishes instance prices by model and performance tier. The checked table lists Rerank 4 Fast and Rerank 4 Pro Medium at $5 per instance-hour, and Rerank 4 Pro Large at $10 per hour. The same reference lists monthly and annual commitments; those figures represent reserved deployment economics rather than the amount charged for one ranked query.

A team running an instance for 200 billable hours at a $5 per-hour rate would have $1,000 of instance-hour costs before any other contractual charges. This example is arithmetic, not an estimate of throughput or achieved searches per second.

Break-even requires actual sustained capacity

A dedicated instance may look cheaper than metered API usage at a certain utilisation, but the threshold cannot be found without documented API prices, traffic shape and measured capacity. Treat an absent rate for a particular model/version as unknown, not zero. Compute a scenario with low, expected and high peak concurrency, plus quiet-hour idle capacity. If demand is highly variable, hourly reservations may leave unused but paid-for capacity.

Pricing comparisons should also account for whether guaranteed performance, private networking or data governance are mandatory. A dedicated deployment may serve organisational reasons unrelated to nominal per-search price.

Chunking makes document length financially relevant

A retrieval engine might return only 40 documents per user query, apparently below an allowance of 100. If some documents are long and the provider splits them into several chunks, the effective work can exceed a straightforward 40-document count. Chunking can affect latency and throughput even where the specific commercial unit depends on your plan. Preserve the size distribution of passages, not just the final result count.

Avoid overoptimising by chopping each source into many tiny fragments; that can increase the number of candidates while hurting the coherence of the evidence supplied to the language model. Evaluation of ranking relevance belongs to a separately documented test.

Differentiate models and billing channels

Rerank 3.5, Rerank 4 and Rerank 5 are different generations and availability may vary between first-party API, Model Vault and cloud partnerships. A per-hour Vault rate from one generation is not evidence of the price for another. The documentation includes model-specific capacity tiers; use its exact identifier, plan and region when issuing a purchase order. Do not infer a public global per-search price from marketing copy when none was captured for the exact API SKU.

A direct API may also have trial or production limits; throughput eligibility is not the same as dollar price. Keep quota metadata separate from financial rows.

Choosing a defensible budget

Start with a measured month of query traffic and candidate-document counts. Estimate model calls, provider chunking and token use in the generation stage, then compare the available commercial options only after attaching official per-SKU rates. The 25 October 2026 review gives verified published Vault amounts but does not provide an immutable launch/effective-date record, a regional contract, or verified per-search Rerank 5 API tariff. Those entries remain null until sourced.