Cohere prices its Command R+ model at $2.50 per million input tokens and $10.00 per million output tokens. This pricing sits between Mistral and Claude Sonnet for the Command line. The company also offers its Embed v4 model in the $0.10 to $0.15 per million token range, which competes with OpenAI text-embedding-3-large. Cohere provides these models through a 1P API, AWS Bedrock, Google Cloud Vertex, and Oracle OCI.
The Command R series provides 50% higher throughput and 20% lower latency than the previous iteration, while halving the hardware footprint. These models support 23 languages and include native tool use and citation capabilities. The Command A model uses a decentralized, expert-driven post-training process where expert models for code, multilingual tasks, reasoning, math, safety, and RAG workflows merge into a single model.
| Model | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
|---|---|---|
| Command R (08-2024) | $0.15 | $0.60 |
| Command R+ (08-2024) | $2.50 | $10.00 |
| Embed v4 | $0.10 – $0.15 | N/A |
Regulated industry constraints and deployment needs
Financial services, healthcare, and government agencies require private deployments via VPC, on-prem, or sovereign cloud. These buyers cannot route sensitive data through OpenAI or Anthropic. Cohere targets these specific needs by focusing on B2B workflows rather than consumer applications. The company recently raised $500 million, bringing its valuation to $5.5 billion, to expand its secure enterprise and sovereign AI solutions.
Healthcare and banking compliance involves strict data privacy and security regulations. HIPAA and the HITECH Act require healthcare providers to protect patient data, while the cost of non-compliance in healthcare averages $9.6 million per organization. Banking compliance costs for retail and corporate banks increased by 60 percent over an eight-year period.
The company maintains a strategy of providing models that work within these constraints without requiring the massive scale of frontier models like GPT-5. Cohere does not train frontier-scale models and instead focuses on models that handle citations and tool use for RAG workflows. This decision means the Command series sits a tier below Gemini Ultra or Claude Opus by design. I find the decision to avoid the frontier race a smart way to maintain higher gross margins, which plausibly sit in the 75 to 85 percent range on direct API sales.
The company uses a variety of strategic cloud partners to deliver these services. Oracle OCI serves as a preferred deployment surface, and the company also works with AWS, Google Cloud, and Azure. If you are managing a migration, you should avoid assuming that replacing a query endpoint will work without rebuilding an index, because a new embedding family requires a new index to avoid degrading retrieval.
RAG reliability and embedding performance
Retrieval-Augmented Generation (RAG) reduces AI hallucinations by connecting a large language model to external, reliable data sources. This architecture allows a model to reference a knowledge base instead of relying solely on its training data. In a RAG system, the process involves collecting information, chunking documents, creating embeddings, retrieving relevant chunks, and generating a response based on that context.
Cohere’s Embed v3 and v4 models serve as the primary retrieval components in many production RAG stacks. These models are cited in the documentation for vector databases like Pinecone, Weaviate, Qdrant, and MongoDB Atlas Vector Search. The Embed 5 family introduces a distinction between Embed 5 Pro, which targets high retrieval quality for offline indexing, and Embed 5 Fast, which targets low latency for interactive search.
| RAG Stage | Primary Driver | Cost Control Method |
|---|---|---|
| Indexing | Embedded content volume | Deduplicate and re-embed changed chunks |
| Query Embedding | Query volume | Reuse identical queries |
| Vector Retrieval | Storage and reads | Tune dimensions and candidate depth |
| Reranking | Billed input tokens | Rerank a focused candidate set |
| Answer Generation | Input/Output tokens | Remove duplicated evidence |
RAG reduces errors but does not eliminate them entirely. A system might retrieve the wrong document if the search engine returns an irrelevant passage, or the LLM might misinterpret the provided context. For example, an AI might see a document stating a plan supports five users and incorrectly tell a user it supports 50.
The company’s Rerank models act as a second-stage scorer for production RAG. Rerank 5 Pro targets quality for difficult enterprise queries, while Rerank 5 Fast targets latency and throughput. Users should budget for the total pipeline cost, including the cost of reranking and long prompts, rather than just the lowest generation rate. Does the reduction in hallucination through specialized reranking justify the added token cost for your specific industry?
