Cohere

Cohere is the AI platform built for companies that cannot send their data to a consumer API. It runs on AWS, Azure and Oracle, inside your own VPC, or entirely on-premises, and its retrieval products — Embed and especially Rerank — are the reason it shows up in serious RAG architectures. Cohere competes on deployment and retrieval, not on chatbot benchmarks.

What is Cohere?

Cohere builds enterprise language models across three product lines. Command handles generation: Command R+ at $2.50 per million input tokens and $10 per million output, Command R at $0.15/$0.60, and Command R7B at $0.0375 per million — the cheap tiers matter because RAG pipelines make enormous numbers of small calls. Embed produces vectors for semantic search, and Rerank reorders retrieved results by relevance, which is consistently the highest-leverage single upgrade to a retrieval system. Above these sit the newer Command A models — A+, A Reasoning, A Translate and A Vision — plus North, an agent platform, and Compass for enterprise search; all of these are sales-led rather than self-serve.

Key Features

• Rerank: The reference product for improving retrieval relevance, and often the quickest win in an underperforming RAG system.
• Deployment flexibility: AWS, Azure, Oracle, your own VPC, or fully on-premises.
• Cheap Command tiers: Command R at $0.15/$0.60 and R7B at $0.0375 per million make high-call-volume pipelines affordable.
• Embed: Multilingual embeddings tuned for enterprise search rather than general similarity.
• North and Compass: Agent platform and enterprise search, sold through sales.

Who Should Use Cohere?

Cohere is for enterprises building retrieval and search over their own data, particularly in finance, healthcare, government and other regulated sectors where on-premises or VPC deployment is a hard requirement. Rerank alone is worth evaluating even if you use another provider for generation. What Cohere is not is a consumer assistant or a frontier reasoning model — for those, OpenAI, Anthropic and Google are ahead. Note also that the newest Command A models require a sales conversation, which slows evaluation compared with self-serve competitors.

Pricing

Command R+ costs $2.50 per million input tokens and $10 per million output. Command R is $0.15/$0.60 and Command R7B $0.0375 per million. Rerank v3 runs around $2 per million. The Command A family, North and Compass are all contact-sales on production keys with no published rates. Enterprise deployments on AWS, Azure, Oracle, VPC or on-premises are quoted individually; on AWS Bedrock, provisioned throughput is about $49.50 per hour per model unit, roughly $29,000 a month. Free trial keys are available for evaluation.

Pros and Cons

Bottom Line

Cohere wins where deployment constraints decide the vendor. If your data cannot leave your infrastructure, it is one of very few credible options, and Rerank makes it worth a look even when it is not your generation provider. Budget for a sales cycle if you want the newest Command A models, and pair it with a frontier API if your product also needs top-tier reasoning.