Research-backed calculator guide

Cohere Rerank and Embed Pricing: Searches, Tokens, Trial Keys and Model Vault

Calculate a Cohere retrieval budget by separating Rerank searches, Embed tokens, trial-key limits, and Model Vault instance commitments.

What the calculator should measure

Start with the workload, not the model name. Record monthly user searches, candidate documents sent to Rerank, tokens embedded during initial indexing, tokens embedded during refreshes, and the chosen deployment route. If you only enter “queries per month,” the result can hide the biggest drivers: one query may rank ten short passages, while another may create many chunks from long documents.

Rerank: convert requests into billable searches

The basic formula is:

Rerank cost = (monthly billable searches ÷ 1,000) × price per 1,000 searches

A billable search is not automatically one document. Under Cohere's definition, it is one query with up to 100 documents to rank.^[1] If your retrieval stage sends 40 candidates for every user query, that is still one Rerank search. If your application makes two independent Rerank calls for one user question, it is two searches. Retries, fallback calls, and offline evaluation passes should be counted when they actually call the paid endpoint.

Long documents need an explicit assumption. Cohere's pricing page says that a document over 500 tokens, including the search query length in its explanation, can be split into multiple chunks, with each chunk counting toward the number of documents ranked for the search.^[1] Before trusting a monthly estimate, measure the post-chunk candidate count in logs. A pipeline that appears to rank 50 source documents may actually send 140 chunks.

Embed: separate backfill from recurring refresh

The basic formula is:

Embed cost = (monthly tokens sent to Embed ÷ 1,000,000) × price per 1M tokens

Include the initial corpus backfill separately from recurring document refreshes. If a 10-million-token corpus is embedded once, that is a one-time 10-million-token event. If 2% of that corpus changes every week, the recurring refresh volume is different from 10 million tokens every week. Also include query embeddings if your search flow embeds each user query. A RAG system often has both document-side and query-side embedding traffic.

Cohere notes that billed units can differ from generic token counts because some tokens are added internally or are model-specific; the billed values are the units charged.^[2] For planning, use the token counts reported by the endpoint or usage export when available. A tokenizer estimate is appropriate for a pre-launch scenario, but label it as an estimate and reconcile it to actual billed usage after launch.

Trial keys: price is zero, capacity is not

A trial-key scenario can show $0 API charge for eligible calls, but it still needs a traffic and rate-limit check. Cohere's documentation distinguishes evaluation keys, which are free but limited, from production keys, which are paid and much less limited.^[4] The same page lists trial and production limits separately for Embed and Rerank. If traffic exceeds the evaluation allowance, a free-cost calculation is not a feasible operating plan.

Use two outputs: estimated cash cost and estimated capacity risk. For example, 120,000 monthly searches over 30 days averages about 4,000 searches per day, but an interactive workload can concentrate most of those calls into a few peak minutes. The monthly average is useful for spend, while requests per minute and burst behavior determine whether a key can serve the traffic.

Model Vault: instance time and commitment

For Standard Vault, calculate the serving commitment independently:

Hourly Vault estimate = instances × active hours × hourly rate

Monthly commitment estimate = instances × published monthly rate

Cohere's Standard Vault documentation says Fixed plans use a fixed number of instances with no autoscaling, while Flex plans can add capacity and bill additional capacity per instance-hour.^[3] It also says rates are per instance and depend on model and performance tier. Therefore, “one model” is not enough input: record the model, tier, instance count, plan type, active hours, baseline, autoscaling ceiling, and commitment term.

A Vault estimate is not automatically comparable to an API estimate on service quality. Vault may buy dedicated capacity, managed infrastructure, or an agreed latency target. If your decision is purely unit price, compare a measured API workload against the instance capacity it would actually replace. If the decision is compliance or isolation, put those requirements beside the arithmetic rather than pretending the cheapest meter is the only criterion.

Worked cost example

Assume a US English RAG application with the following monthly workload. The 120,000 searches are paid-path Rerank calls. Each call sends no more than 100 candidate chunks after retrieval. The 12 million Embed tokens include document refreshes and query embeddings. The prices below are an illustrative Azure-hosted route from the published Azure Foundry Cohere table, not a direct promise for every Cohere account or cloud deployment.^[5]

The hosted API subtotal is deliberately narrow. It excludes the vector database, generation model, application servers, network egress, logging, taxes, support plans, and engineering time. It also excludes a separate reranker instance because the example assumes Rerank is consumed through the hosted API route. If you instead deploy both Embed and Rerank in Vault, price both instance selections, account for their performance tiers, and model baseline plus burst capacity under the selected plan.^[3]

The table also demonstrates why the meters cannot be merged into one “tokens per month” figure. Rerank price is driven by search count, not by a simple Embed-token conversion. Embed price is driven by tokens, while Vault price is driven by instance time or commitment. A single total is useful only after each component has been normalized in its own unit.

For a sensitivity check, double only one input at a time. At 240,000 searches, the illustrative Rerank line becomes $480 while Embed remains $1.44. At 24 million Embed tokens, the Embed line becomes $2.88 while Rerank remains $240. If the Rerank candidate set crosses chunking thresholds, keep the search count assumption explicit and inspect the endpoint's billing behavior rather than assuming that more chunks create a new search automatically.

How to make the estimate auditable

Store the input snapshot with the date checked, endpoint or cloud route, model version, region, currency, and whether the figure is a vendor quote or an estimate. Current pricing can change, and the same Cohere model family may appear through Cohere's API, a cloud marketplace, or a private deployment with different commercial terms. The Azure page is useful for a concrete hosted comparison, but it is not a substitute for the contract or account-specific rate card.^[5]

Then reconcile estimates to usage. Export monthly counts for Rerank requests, Embed tokens, retries, and failed calls. Compare those counts with invoice line items. If the estimate is lower than the invoice, investigate duplicate retries, chunking, refresh jobs, multiple regions, or a different model tier before changing the unit-rate assumption. The AI API invoice reconciliation calculator is useful for mapping an estimate to billed rows, while the AI token cost calculator can help with token-volume scenarios.

For a full RAG budget, add the generation and storage layers after this page's Cohere retrieval calculation. The RAG cost calculator helps place retrieval, embeddings, vector storage, and generation in one workflow view. If you are comparing different vendors or routes, use the model pricing comparison calculator but preserve each provider's native billing unit until the final total.

FAQ

Is Cohere Rerank priced per search or per document?

Cohere documents Rerank pricing by search. Its definition is one query with up to 100 documents to be ranked.^[1] A request that ranks 40 candidates is not four-tenths of a search. However, long documents may be split into chunks, and those chunks count toward the number of documents ranked. Count API calls and inspect chunking behavior rather than converting document count directly into price.

What is the Cohere Embed API price per token?

There is no single rate that should be copied across every route and model version. Cohere's documentation establishes the meter—tokens embedded—while the Azure Foundry Cohere listing provides a concrete hosted example of $0.12 per 1 million text tokens for Embed v4 Global as checked on 2026-11-02.^{[2] [5]} Use that figure only for the stated Azure route and replace it when your account uses another model, region, or commercial agreement.

Are Cohere trial keys free for production traffic?

No. Cohere says trial API usage is free but limited, and trial keys are not the same as a production billing arrangement.^{[1] [2]} The rate-limit documentation records a monthly call limitation and separate endpoint limits for evaluation and production keys.^[4] A trial can support experiments and a proof of concept; it should not be used to claim that a production workload will cost zero or fit within the same capacity.

When is Model Vault cheaper than API usage?

Model Vault is cheaper only under the workload, capacity, and commercial assumptions you supply. A published hourly rate can look attractive at low active hours but exceed a monthly commitment at sustained utilization, while a monthly instance may provide capacity or isolation that a pay-as-you-go API does not. Compare the API workload cost with the number and tier of Vault instances needed, and include idle baseline, burst capacity, SLO, and operational exclusions.^[3]

Should I add Model Vault charges to Rerank and Embed API charges?

Not when Vault is replacing those hosted calls for the same path. Treat API and Vault as alternative deployment lanes unless your architecture explicitly uses both. You might pay for a Vault embedder while also calling a hosted Rerank API, in which case include one line for each actual service. Write the architecture beside the calculation so that a deployment choice is not double-counted.

How should I budget indexing and query embeddings?

Separate one-time corpus indexing, scheduled refreshes, and per-query embeddings. For each lane, multiply tokens by the applicable per-million-token rate and record the date and model. If token counts are estimates, state the tokenizer and average document/query assumptions. After launch, replace estimates with billed usage or endpoint usage exports and reconcile the difference.

Does the calculator include vector database or generation costs?

No. The worked example covers only the stated Rerank and Embed lines and compares a separate Model Vault alternative. Vector storage, retrieval infrastructure, LLM answer generation, network, observability, support, tax, and engineering labor are exclusions unless you add them as separate inputs. This separation keeps the Cohere retrieval calculation inspectable and prevents an apparently precise total from hiding omitted services.