Cohere’s useful pricing comparison is the entire retrieval pipeline, not just the chatbot. Command generates the answer, Embed builds searchable vectors, and Rerank reorders retrieved evidence. Embed 5 adds a practical choice between quality-focused indexing and latency-focused queries.
The table above renders our maintained Cohere pricing dataset at build time. It is a direct API reference, not an enterprise contract quote. Embedding rows measure ingestion rather than answer generation; a zero output rate does not mean embedding is free. The chart and mini-calculator compare Command generation only.
Which Cohere models should you test?
| Workload | Starting candidate | What to measure |
|---|---|---|
| Classification, routing, short grounded answers | Command R7B | Valid outputs, retries, accepted-answer cost |
| Enterprise RAG and tool use | Command A | Citation accuracy, task success, latency |
| Quality-critical corpus indexing | Embed 5 Pro | Recall on your documents, indexing cost |
| Interactive search and agent queries | Embed 5 Fast | Recall against the chosen index, p95 latency |
| Existing Embed v3 or Embed 4 integration | Current endpoint plus a shadow Embed 5 index | Migration benefit and cutover risk |
| Ranking retrieved candidates | Rerank 5 Pro or Fast | Ranking lift, billed input tokens, document limits |
| Pricing data: How we verify → | ||
Command R7B’s retained rates are historical observations, not an established current low-cost offer. Obtain current exact rates before a production comparison. Command A belongs in the same evaluation when a narrow model needs repeated attempts or fails to ground answers reliably.
The R and R+ rows are marked legacy in our dataset. Existing integrations should verify their pinned endpoint’s lifecycle; new projects should not assume a familiar name guarantees long-term availability. Specialist Command variants and enterprise routes also need an access check.
Embed 5 Pro versus Fast: the retrieval decision
Cohere’s official release notes identify embed-v5.0-pro and embed-v5.0-fast. Pro targets highest retrieval quality, especially offline indexing; Fast targets low latency and high throughput for interactive search and agent loops.
Pro and Fast share an embedding space. Cohere recommends indexing documents with Pro and embedding queries with Fast. That compatibility is specific to this pair: it does not establish compatibility with an existing Embed v3 or Embed 4 index.
The family supports text, images, mixed document inputs, more than 100 languages, and a 128k-token context window. Output dimensions range from 256 to 2,048, with float, int8, and binary options. Smaller vectors can reduce storage, but test recall before treating compression as a saving.
Compare three configurations on identical queries: all-Fast, Pro-index/Fast-query, and your current model. Read our Embed 5 launch coverage for additional deployment context.
Rerank 5 Pro versus Fast: token billing
The live table now includes rerank-v5.0-pro and rerank-v5.0-fast. Pro targets quality on difficult enterprise queries; Fast targets latency and throughput. Both support a 32,768-token query-plus-document context and optional Context Focus, which returns relevant passages in the same call.
Cohere’s current pricing cards and Rerank 5 launch article agree on token-based API prices. The v5 API example reports billed_units.input_tokens. Use that billed total, not a generic search count or an assumed fixed document size. The older search-unit FAQ is not the Rerank 5 rate card.
Compare Pro and Fast using the maintained rates. Model Vault remains a separate per-instance deployment product; its hourly and monthly charges are not shared API prices.
How to budget a Cohere RAG system
| Stage | Billing or infrastructure driver | Main cost control |
|---|---|---|
| Indexing | Embedded content and refresh frequency | Deduplicate; re-embed only changed chunks |
| Query embedding | Query volume and chosen model | Reuse identical queries where appropriate |
| Vector retrieval | Storage, reads, replicas | Tune dimensions and candidate depth |
| Reranking | Rerank 5 billed input tokens; older endpoints may use search units | Rerank a focused candidate set |
| Answer generation | Billed input and output tokens | Remove duplicated evidence; cap output |
| Evaluation and retries | Repeated pipeline runs and review | Limit retries; track accepted answers |
| Pricing data: How we verify → | ||
Budget indexing and query embedding separately. For a text-only workload, multiply each stage’s token volume by its corresponding live rate; then add vector infrastructure, reranking, and generation. Image-heavy inputs require the provider’s modality-specific accounting.
For example, a corpus with ten million text tokens and one million monthly query tokens has a one-time indexing workload plus recurring query traffic. Do not add generated answer tokens to the embedding estimate. Use the token calculator separately for Command’s input/output mix.
The decision metric is total pipeline cost divided by accepted grounded answers. A lower embedding rate can lose if retrieval errors force more reranking, longer prompts, or repeated generation.
Access, deployment, and hidden costs
Evaluation access is not a production capacity commitment. Confirm the selected model, key type, rate limits, and support route before promising throughput. Newer or specialist endpoints may require sales contact.
Cohere offers direct API and managed deployment options. Its public pricing page also lists per-instance deployment charges: those are not interchangeable with token-metered API rates. Model Vault, Microsoft Foundry, and Amazon SageMaker require route-specific pricing and capacity checks.
During migration, include parallel-index storage, vector writes, engineering time, and rollback capacity. A new embedding family generally needs a rebuilt index; replacing the query endpoint alone can silently degrade retrieval.
Before cutover, compare recall at k, citation correctness, accepted-answer cost, and latency on representative multilingual and document-heavy queries. Keep the existing index until the new route passes your quality threshold.
Cohere versus alternative stacks
Cohere is worth shortlisting when enterprise retrieval, multilingual documents, or private deployment are central requirements. Compare the entire Embed-Rerank-Command stack, not unrelated models on a single token-price leaderboard.
OpenAI pricing provides another generation and embedding baseline. Anthropic pricing belongs in answer-quality tests for long-document and agent workloads. Our best AI for RAG guide explains how to compare these roles separately.
If managed open models fit your generation requirements, benchmark Novita’s hosted endpoints as an alternative generator—not as a drop-in replacement for Cohere’s embedding space.
Affiliate disclosure: we may earn a commission from the link above, at no extra cost to you. Compensation does not affect our data or recommendations.
Rerank 5 pricing and billing checked October 22, 2026 against the pricing page, launch article and API reference. Earlier Embed review, October 13, 2026: official Embed 5 release notes and Cohere pricing. The table uses Data-maintained canonical rates; confirm your contracted route before budgeting production.
FAQ
Should I index with Embed 5 Pro and query with Fast?
Cohere recommends that combination because the two models share an embedding space. Compare it with all-Fast on your own retrieval evaluation before migrating.
Can Embed 5 query my existing Embed v3 index?
Do not assume cross-family compatibility. Build and evaluate a separate index before replacing the production query route.
Is Rerank billed like Command?
Rerank 5 is billed per input token, with no generated output-token charge; Command has input and output token rates. Older Rerank endpoints may use search units. Do not convert search units or Vault instance charges into a token rate.
Is the lowest generation rate the cheapest RAG system?
Not necessarily. Retrieval quality, reranking, prompt length, and retries determine cost per accepted answer.