Cohere

Cohere is a Canadian enterprise AI company that sells Command language models, Embed and Rerank retrieval models and the North agent platform, through an API and through private deployments. Free trial keys allow 20 chat requests a minute. Its security page says you can opt out of training, and it offers VPC, on-premises and Model Vault hosting.

Facts at a glance

Prices checked 2026-10-31: no per-token prices were visible on the pages we read, so this entry has no price table. The model entry is Command models.

Models offered

Cohere's generative line is called Command. The retrieval line, Embed, Rerank and the document parser, is what many teams actually buy from Cohere.

Command A+ is a mixture-of-experts model with 218 billion total and 25 billion active parameters. Cohere describes it as the last model in the Command A family, unifying reasoning, vision, translation and tool use that were split across earlier models, and it covers 48 languages against 23 for Command A. The early-2024 Command R generation was deprecated on 15 September 2025. Cohere also lists its retrieval and speech models on the pricing page: Embed, Rerank, Parse, Transcribe and a translation model.

The product that sits on top is North, described as an all-in-one AI platform for workplace tasks with agents, and Compass for search and discovery. Those are products, not APIs, and Cohere prices them through sales.

Rate limits and keys

Cohere has two kinds of API key: evaluation keys, which are free but limited, and production keys, which are paid and much less limited. The rate-limit page lists per-model limits.

Trial keys, and production keys on the newer chat variants, are capped at 1,000 API calls a month. That is a real ceiling for a prototype, so a team evaluating Command A+ should expect to talk to sales early, because even a production key on the newest models needs a conversation. The practical lesson is that Cohere's self-serve path is for older models and retrieval, while its newest generative models go through a sales process.

Data retention and training

The security page makes three statements that matter to a buyer.

• Control. "Maintain full control over your data with customizable handling and retention settings. Opt out of model training at any time."
• Compliance. The API platform is hosted on major cloud infrastructure and is SOC 2 Type II compliant, with annual third-party audits and penetration tests. Cohere also says it tracks the lineage of its training data to guard against data poisoning.
• Deployment. You can bring the models to your own infrastructure through a virtual private cloud or on-premises set-up, or use Model Vault, a dedicated, fully managed platform with isolated inference and guaranteed performance.

What we could not find is the default: whether the hosted API trains on inputs unless you opt out, how long prompts are kept, and which regions host the platform. The page speaks of opt-out, which implies the default may allow training on some data. For a regulated company the sensible route is the one Cohere itself pushes: private deployment, where the data never reaches Cohere's cloud.

Private deployment and open weights

Cohere's differentiator is that it sells a way out of its own cloud.

Custom pricing for the private routes is set by enterprise needs, and Cohere also offers model customization to enterprise customers. The Command A+ model card lists example minimum GPUs: a 4-bit build on one B200 or two H100s, an 8-bit build on two B200s or four H100s, and a 16-bit build on four B200s or eight H100s. Cohere says the quality differences between the three builds are negligible and recommends the 4-bit one for most uses. This is a clear hardware table by the standards of model cards, and it tells you a private deployment is a GPU server project, not a laptop job. See GPU requirements for LLMs and private AI for teams.

The older Command A weights are CC-BY-NC 4.0, which bars commercial use. Apache 2.0 for Command A+ is a clear break from that habit. Read the licence of the exact weights you plan to use.

Sizing a private deployment

A private deployment sounds simple until the hardware quote arrives. The numbers below are our arithmetic from the model card's parameter count, and they are rough.

The weights are not the whole footprint. Context cache and batch size add memory on top, which is why the card lists more GPUs than the weight size alone would suggest. Because Command A+ activates only 25 billion parameters per token, it generates faster than a dense model of the same total size, but it still has to hold all 218 billion values in memory. If you plan an on-premises deployment, ask Cohere or your integrator for a throughput figure at your expected concurrency, and do not rely on the card's minimum GPU list for production sizing. For a managed route, Private AI setup services explains what a setup project includes.

Questions to ask before you sign

Cohere sells through sales conversations, so the questions matter more than a price table would. Take this list to the first call.

Write the answers into the contract. Statements on a marketing page can change, and a clause cannot be edited without your signature.

Retrieval is the quiet half of Cohere

Many Cohere customers use it for search rather than chat. The pricing page lists Embed, Rerank and Parse as new or updated products, and it announces Rerank 5 with a Context Focus feature that aims for sharper results and fewer tokens. The rate-limit table shows what that means in practice: Embed allows 2,000 inputs a minute even on trial keys, and Rerank allows 1,000 requests a minute on production. A retrieval-augmented agent calls Embed to index, Rerank to pick the best passages and Command to write the answer, so the limits that matter to you are three, not one. For the idea behind that pattern, see MCP vs RAG.

If your project is a search or support agent, the sensible first test is cheap. Build the pipeline on trial keys, measure retrieval quality on your own documents, then ask sales about production limits only if the quality is good enough. That order avoids a negotiation before you know the product fits.

Moving to or from Cohere

1. Start with retrieval. Embed and Rerank are the easiest pieces to swap in. Vectors from one model do not mix with vectors from another, so plan a full re-index when you change the embedding model.
2. Chat format. Cohere's v2 chat API has its own shape. Use a compatibility endpoint if you want an OpenAI-style client.
3. Model names. API names carry dates, for example command-a-plus-05-2026. Pin the exact id so a new snapshot does not change behaviour under you.
4. Limits. Budget for the 1,000 calls a month on trial keys, and ask for production keys before any pilot with real users.
5. Exit. If you self-host Command A+ weights, your exit is easy because the Apache 2.0 licence lets you keep them. If you use the hosted API or Model Vault, plan for re-indexing and prompt re-tuning when you leave.

Which agents can use it, and how

Cohere's chat API has its own format, and it also offers compatibility endpoints for OpenAI-style clients. For a coding agent, the picture is thin. We did not find Cohere in the provider lists of Claude Code, Codex CLI or Gemini CLI. Agents with a custom OpenAI-compatible provider option, such as OpenCode and Cline, can in principle point at an OpenAI-style endpoint. Cohere also ships North Mini Code, a coding model listed in its rate limit table at 500 requests a minute in production.

If you self-host Command A+ weights with a server such as vLLM, any agent that accepts a custom endpoint can use them. See vLLM and Ollama for serving routes. Cohere's strengths are retrieval and enterprise search, so the natural fit is a RAG or search agent built on Embed, Rerank and Command, not a coding agent.

Alternatives

Who should pick Cohere

A pre-launch checklist

1. Ask sales for per-token prices, since the pricing page leads with custom enterprise terms.
2. Get production limits for Command A+ in writing before you design for volume.
3. Opt out of training in your account settings and record it.
4. Decide between hosted, Model Vault and private deployment on your data rule.
5. If self-hosting Command A+, pick the 4-bit build and size GPUs from the model card.
6. Check the licence of each weight set you use.
7. Request the SOC 2 Type II report and DPA for your security review.

Our take

Cohere is a company for buyers who want control: opt-out of training, SOC 2 Type II, private deployment in a VPC or on-premises, and now Apache 2.0 weights for its newest model. It is not a company for someone who wants to sign up, read a price table and ship. The pricing is custom, the newest models need a sales conversation for production limits, and the default retention and region terms were not on the pages we read. If your constraint is "no data leaves the building" and you can afford GPU hardware or a managed vault, it is a serious option. If your constraint is price per token, look at the open-weight hosts.

We have not run tests for this page. It reflects Cohere's published pages on 2026-10-31. Cohere does not pay for placement here.

Related pages

Command models, Private AI for teams, self-hosted AI, LLM API prices, Enterprise AI agents.