Cohere Rerank Pricing for RAG Teams

Cohere Rerank pricing matters when search quality is the part of your RAG stack customers can feel. A weak retrieval layer hands the model mediocre context, then charges you to write a polished answer around it. Rerank sits between retrieval and generation, where it can sort a broad candidate set into the documents worth sending forward.

Cohere’s product line splits that work across Command for generation, Embed for vector representations, and Rerank for ordering search results. The useful question is not which model has the most impressive sounding name. It’s where each one belongs in the request path, and which cost unit you’ll be paying as usage grows.

TLDR

Cohere Rerank pricing belongs in the retrieval budget, while Embed handles indexing and Command writes the answer. Cohere’s free trial allows 100 API calls per minute and 1,000 calls per month. That is enough to test a workflow before production traffic arrives.

Cohere Rerank Pricing at a Glance

Rerank is built for the moment after your search system has produced a set of possible documents and before your generation model sees them. It evaluates the query against those candidates and pushes the most relevant material to the top.

That makes its price worth evaluating differently from a generation model’s token bill. A Rerank request is part of the cost of finding evidence. Command is part of the cost of turning that evidence into a response. Embed is part of the cost of preparing and retrieving the underlying corpus.

If your application answers questions over internal documents, product records, support articles, or long reports, the retrieval path usually deserves more scrutiny than it gets. Teams often obsess over the final answer model because it is visible. Users experience the retrieval failure first.

A RAG application can return a fluent answer that is wrong for a dull reason: the relevant document never made it into the model’s context. Reranking gives you another filter between broad recall and final generation.

Cohere’s free trial allows 100 API calls per minute Related analysis. That rate is useful for testing how a reranking step behaves under short bursts of evaluation traffic, including repeated searches against the same set of documents.

Cohere’s free trial allows 1,000 calls per month Related analysis. That is a prototyping allowance, not a production budget, but it gives developers room to compare result quality before a feature becomes part of every customer request.

The pricing decision comes down to the job you assign each service. Rerank should earn its place by improving the documents that reach the answer model. Embed should earn its place by making retrieval fast and relevant. Command should earn its place by producing an answer that follows the evidence rather than wandering off into the weeds.

Command, Embed, and Rerank Use Cases

Command is the generation layer. It reads the selected context, follows instructions, synthesizes information, and produces the final response. Use it when the application needs a written answer, classification with explanation, extraction, summarization, or an agent-style action plan.

Command R+ is listed at $2.50 input and $10 output per 1M tokens Related analysis. Its price profile fits work where answer quality, longer context handling, and complex reasoning justify a higher generation bill.

Command R+ has a 128K context window Related analysis. A larger window can help when the application needs to inspect lengthy source material, though a strong retrieval path still keeps unnecessary text out of the prompt.

Command R is listed at $0.15 input and $0.60 output per 1M tokens Related analysis. It gives teams a lower-cost Command option for workloads where the answer format is narrower, the selected context is cleaner, or usage volume makes every generated token more consequential.

Embed is the representation layer. It turns text into vectors so your application can retrieve semantically related material instead of relying only on literal keyword matches. It belongs at ingestion time, when documents enter the system, and at query time, when a user’s question needs a comparable representation.

Embed v3 is listed at $0.10 per 1M tokens Related analysis. That makes embedding cost easier to separate from the request path: one part comes from preparing source material, while another comes from embedding each incoming query.

Rerank is the judgment layer for retrieval results. A vector database may return material that is broadly related to the query. Broadly related is often not enough. A support assistant needs the policy that answers the customer’s question, not a pile of documents that share the same nouns.

This is why the three products work best as parts of one pipeline. Embed finds candidates. Rerank sorts candidates by query relevance. Command uses the selected material to produce the response.

Product	Pricing basis available here	Context window	Best fit
Command R+	$2.50 input and $10 output per 1M tokens	128K	Higher-stakes generation and larger source sets
Command R	$0.15 input and $0.60 output per 1M tokens	128K	Cost-sensitive generation with focused context
Embed v3	$0.10 per 1M tokens	n/a	Indexing and semantic retrieval
Rerank	Per-search evaluation	n/a	Ordering retrieved candidates before generation

The table is a model-selection tool, not a license to treat model choice as an architecture. The best RAG systems reduce the amount of context that needs to be generated over in the first place. Clean document chunking, sensible metadata, and disciplined retrieval make the lower-cost generation option more viable.

A RAG glossary is useful alongside the $0.10 per 1M tokens Embed v3 price because indexing and retrieval vocabulary gets muddy fast once a project includes vectors, chunks, candidates, and reranking. Those are separate jobs with separate costs.

Free Trial Limits for Evaluation

The free trial is best used to answer a narrow question: does Rerank improve the evidence your application sends to Command?

Start with a representative collection of documents. Use the material your customers will search, including the messy cases: near-duplicates, outdated versions, documents with similar titles, and long pages where the answer is buried halfway down. A tidy demo corpus tells you very little.

Create a set of queries that reflect how people ask for help. Include direct requests, vague requests, and queries where a keyword-only system would likely fetch the wrong document. Then inspect the results before and after the reranking step.

Cohere’s free trial allows 100 API calls per minute Related analysis. That is enough headroom for an evaluation script that runs query variants, compares candidate ordering, and captures examples for human review.

Cohere’s free trial allows 1,000 calls per month Related analysis. Spend those calls on meaningful comparisons instead of burning them on a polished demo that only proves your most obvious documents can be found.

The core evaluation should focus on relevance at the top of the list. Did the right document move upward? Did misleading but semantically adjacent documents fall away? Does the final answer cite or reflect the material a domain expert would choose?

Keep generation out of the first pass if you can. Evaluate retrieval and reranking independently, then bring Command into the workflow after you understand which source material reaches it. Otherwise, a persuasive answer can hide a retrieval problem.

There is another practical benefit to testing this way. You can decide whether reranking belongs on every query or only on requests where recall is broad, stakes are higher, or users are searching a particularly large corpus. The product decision comes from observed result quality, not a generic belief that more model calls always produce better answers.

Cost Choices for a RAG Workflow

A RAG workflow has different cost centers, and mixing them together makes budgeting harder than it needs to be.

Embedding costs occur when you prepare source material and when you encode incoming queries. Generation costs depend on prompt size and answer length. Reranking costs track the search evaluation layer. Each can be tuned without rewriting the whole system.

Embed v3 is listed at $0.10 per 1M tokens Related analysis. That figure belongs in the ingestion plan, especially when you are estimating the cost of adding a large body of documents or reprocessing content after a chunking change.

Command R is listed at $0.15 input and $0.60 output per 1M tokens Related analysis. Its lower token pricing changes the economics of applications that generate many concise responses from tightly selected context.

Command R+ is listed at $2.50 input and $10 output per 1M tokens Related analysis. The gap between the Command options gives you a reason to route work by task difficulty rather than running every request through the more expensive model.

That routing can be simple. Use the lower-cost model for repeatable requests where retrieval is strong and response expectations are constrained. Reserve the higher-priced option for requests involving lengthy source material, more complex synthesis, or outputs where a weak answer creates a customer problem.

Command R+ has a 128K context window Related analysis. Context capacity gives developers flexibility, but dumping an entire document set into a prompt is still a bad habit. Bigger prompts create bigger bills and can dilute the evidence the model should prioritize.

A prompt caching glossary pairs naturally with the $2.50 input and $10 output per 1M tokens Command R+ price because repeated instructions and stable context can become a meaningful part of a generation budget. The bill is shaped by what you send repeatedly, rather than only by the answer users see.

Reranking creates a different tradeoff. It adds another request to the path, but it can reduce the amount of low-value context sent to Command. That tradeoff is worth testing with your own documents. If reranking removes irrelevant text and improves answer grounding, it may lower generation waste while making the product better.

The winning architecture is usually the one that spends money closest to the failure it is fixing. Use Embed to retrieve. Use Rerank when candidate ordering is the weak point. Use Command to write the answer once the relevant material is in hand.

A search assistant with cheap generation and sloppy retrieval is a literal money pit. It produces plenty of text, then makes users sort out whether any of it is based on the right source.

How to Choose a Cohere Setup

Start from the user experience you are trying to build.

For a knowledge-base assistant, the retrieval layer often decides whether users trust the product. Embed creates the search index, Rerank improves which passages rise to the top, and Command turns those passages into a clear response. Test the retrieval quality before debating answer style.

For document processing, Command may carry more of the workload. If the task is to extract, summarize, classify, or transform known input, you may not need a full retrieval workflow for every request. Embed and Rerank become more valuable once users need to search across a changing document collection.

For an internal research tool, context size and source selection pull in opposite directions. Command R+ has a 128K context window Related analysis, but a careful Rerank stage can still keep the final prompt focused on the source passages that answer the question.

For an early prototype, the free trial sets the boundaries. Cohere’s free trial allows 100 API calls per minute Related analysis and 1,000 calls per month Related analysis. Use that period to inspect relevance, measure response consistency, and identify which part of the workflow causes the most expensive failures.

The product line gives you enough separation to make deliberate choices. Generation quality is one choice. Retrieval quality is another. Reranking is where the search layer gets a chance to prove it found the right material before the model starts talking.

Key Takeaways

• Rerank belongs between retrieval and generation, where it can improve the candidate documents sent to Command.
• Command R+ is listed at $2.50 input and $10 output per 1M tokens, while Command R is listed at $0.15 input and $0.60 output per 1M tokens.
• Embed v3 is listed at $0.10 per 1M tokens for semantic retrieval work.
• Cohere’s free trial allows 100 API calls per minute and 1,000 calls per month for evaluation.
• A narrower, better-ranked context set can improve answers and control generation spend.

Sources

• PE Collective Cohere pricing guide, 100 API calls per minute
• PE Collective Cohere pricing guide, 1,000 calls per month
• PE Collective Cohere pricing guide, $2.50 input and $10 output per 1M tokens
• PE Collective Cohere pricing guide, $0.10 per 1M tokens