Cohere is easiest to evaluate as a retrieval stack, not as a single chatbot API. Command generates grounded answers, Embed turns content into searchable vectors, and Rerank improves the ordering of retrieved results before generation.
The live modules above pull tracked Command rates from our daily dataset. Use the Cohere pricing page for the full model list and update history, or enter your own token mix in the AI token calculator.
Which Cohere Model Should You Use?
| Workload | Start with | Escalate to | Measure |
|---|---|---|---|
| Classification, routing, short retrieval answers | Command R7B | Command R | Valid outputs and retry rate |
| Enterprise RAG and grounded chat | Command A | Command A+ or a managed deployment | Answer quality with citations |
| Long-document generation on a stable endpoint | Command R | Command A | Cost per accepted answer |
| Existing R+ production workload | Pinned Command R+ endpoint | Command A migration test | Quality, latency, and migration effort |
| Multilingual or multimodal retrieval | Embed 4 plus Rerank 4 | Command A family | Recall, ranking quality, and answer success |
Command R7B is the economical first test for narrow tasks. Its smaller footprint also makes it useful when deployment efficiency matters, but low token cost does not compensate for extra retries or weak retrieval.
Command A is the stronger general candidate for tool use, agents, multilingual work, and retrieval-augmented generation. Cohere documents a larger context window for Command A than for the R-series endpoints, but context capacity should be treated as a ceiling rather than a target.
Command R and Command R+ remain relevant to pinned production integrations. New projects should compare them with the Command A family and confirm the exact endpoint status before committing to a long migration.
How Cohere Pricing Works
Cohere bills Command generation by input and output tokens. The Chat API returns both total tokens and billed_units; the billed values are the ones used for the invoice when Cohere adds internal tokens that the customer did not control.
Embed is billed by the amount of content embedded, while Rerank is billed by searches. These units should not be folded into one imaginary per-token rate:
| Pipeline stage | Billing driver | Main cost control |
|---|---|---|
| Document ingestion | Embedded text, images, and refresh frequency | Deduplicate content and update only changed chunks |
| Retrieval | Vector database reads and storage | Limit candidate count and unnecessary replicas |
| Reranking | Search volume and documents sent per query | Rerank a focused candidate set |
| Answer generation | Billed input and output tokens | Retrieve less, cap output, and reuse stable context |
| Evaluation | Test traffic and human review | Use representative samples and automated checks |
A cheap generation model can still produce an expensive RAG system if every request retrieves too many chunks, reranks a large candidate set, and sends duplicated evidence into the prompt.
Current Model and Access Notes
Cohere’s current documentation lists Command A+, Command A, Command R7B, specialist Command A variants, Command R+, and Command R. It also lists Embed 4 for text, image, and mixed-document retrieval, plus Rerank 4 Fast and Rerank 4 Pro.
Not every model follows the same self-serve production path. Cohere’s rate-limit documentation says standard Command endpoints receive higher production limits, while newer variants such as Command A+ and Command A Reasoning require sales contact for production access. Confirm availability in the dashboard before designing around an evaluation endpoint.
Trial keys are free but limited. They are useful for a proof of concept, not for estimating production throughput. Test the production-key limit, regional deployment option, latency target, and support path alongside model quality.
Review the official Cohere pricing, model catalog, and rate-limit documentation before launch. Marketplace pricing on AWS, Azure, or Oracle can differ from Cohere’s direct route.
Hidden Costs in a Cohere RAG Stack
Index refreshes can dominate embedding cost. Re-embedding an entire corpus after small edits wastes both API calls and vector-database writes. Use stable document identifiers, content hashes, and incremental updates.
Rerank quality depends on candidate quality. A reranker cannot recover documents that retrieval never returned. Track recall before reranking, then test whether the extra search charge improves successful answers enough to pay for itself.
Long context can hide retrieval mistakes. Sending more chunks may improve recall while lowering precision, increasing prompt cost, and distracting the generator. Tune chunk size, overlap, candidate count, and rerank cutoff together.
Retries multiply the whole pipeline. One failed answer may trigger another vector search, rerank request, generation call, and validation pass. Budget at the task level and cap retries.
Private deployment changes the cost model. Cohere offers managed and enterprise deployment options where instance commitments, capacity, support, and data controls matter more than public API token rates. Compare total annual cost and utilization, not only the model card.
Cohere vs OpenAI, Claude, and Managed Open Models
Cohere is a strong shortlist choice when retrieval quality, multilingual enterprise search, private deployment, and a unified Embed-Rerank-Command stack matter. The narrower product focus can simplify ownership for RAG teams.
OpenAI pricing is the broader comparison for general reasoning, tool ecosystems, and multimodal applications. Anthropic pricing belongs in tests centered on long documents, coding, and agent work. For the market-wide view, use our AI API pricing comparison.
If an open-model route fits your data and quality requirements, benchmark Novita’s managed endpoints against the direct Cohere stack.
Affiliate disclosure: we may earn a commission if you use the link above, at no extra cost to you. Compensation does not affect our pricing data or recommendations.
FAQ
Does Cohere offer a free API tier?
Cohere provides free evaluation keys with lower rate limits and a monthly request cap. Production keys are paid and have higher limits for standard endpoints.
Is Command R7B the cheapest Cohere generation model?
It is the lowest-priced active Command model in our current tracked generation data. Validate task success and retry volume before choosing it for production.
Are Embed and Rerank charged like Command?
No. Command is token-metered for generation, Embed is based on embedded content, and Rerank is based on searches.
Which Cohere model is best for RAG?
Start by testing Command A with Cohere’s current Embed and Rerank models. For high-volume narrow workloads, compare Command R7B or Command R on cost per successful grounded answer.
Can Cohere run in a private enterprise environment?
Cohere offers direct API access plus managed and enterprise deployment options across major cloud platforms. Availability, capacity, and contract terms depend on the selected model and platform.
Bottom Line
Cohere makes the most sense when retrieval is the product. Start with Command R7B for economical narrow work or Command A for the general enterprise route, then measure the complete Embed-Rerank-Command pipeline. The winning design is the one with the lowest cost per grounded, accepted answer—not the lowest generation rate in isolation.