Cohere: our review
Cohere is a Canadian language-model vendor built for enterprises, with a generation, search, and reranking API billed per token. It's a platform for developers who want a North American enterprise contract, not a consumer chatbot.
What we take away from it
What counts in its favor
- Cohere states plainly that it doesn't intentionally train on data sent through the API, and Enterprise accounts can negotiate a data processing agreement (DPA) to govern their use.
- The cheapest model in the current lineup, Command R7B, charges $0.0375 per million input tokens and $0.15 per million output tokens — well below the entry price of the previous generation, Command R+, at $2.50 and $10.
- Cohere publishes a publicly browsable subprocessor list (trustcenter.cohere.com/subprocessors) and claims SOC 2 Type II certification on its API infrastructure.
What counts against it
- The free trial key is explicitly banned from production or commercial use in the documentation: it's not a real free tier, just a test before paying.
- No fixed processing region is guaranteed by default: the privacy policy cites transfers to Canada, the United States, and the United Kingdom, and the API infrastructure runs on Google Cloud in the United States.
- For trial and research accounts, the documentation warns that inputs and outputs may be used for research work: protection against training fully applies only to a paying Enterprise account.
A sound choice for a technical team that wants a North American enterprise contract with negotiable guarantees on data, but not an obvious pick for anyone seeking proven European hosting by default.
The facts
- Data hosting
- Cloud or self-hosted · source
- Categories
- Platforms and APIs,Language models
- Data verified on
- 30 July 2026
What happens to your data
- Is your data used to train their models?
- Depends on the plan, or an option you can turn off
- Subprocessors
- Public list at trustcenter.cohere.com/subprocessors: API hosting on Google Cloud (United States) and various tools (Fullstory, LaunchDarkly, New Relic, Retool, Sentry, Segment, Sendgrid, Vercel), all located in the United States per the page reviewed. The privacy policy also mentions data transfers to Canada and the United Kingdom.
- Retention
- Inputs and outputs are kept for about 30 days for Enterprise accounts, per the privacy policy; deletion on request at privacy@cohere.com for other account types.
Cohere in detail
Cohere is a Canadian company founded by former authors of the paper that launched Transformer models. It sells a language-model API — generation, search (Embed), ranking (Rerank) — with a stated positioning: serving businesses that want a contract and guarantees, not one more consumer chatbot.
A model builder, not just a reseller
Cohere trains its own models (the Command family) rather than reselling access to a third party's, which sets it apart from a router like OpenRouter. The offer covers three uses: Command for text generation, Embed for turning a document into a search vector, Rerank for ranking results by relevance. A business building an internal search engine or a document-sorting tool finds a catalogue built for that job, more than for conversation.
Pricing, generation by generation
The pricing grid clearly separates model generations. Command R7B, the newest and cheapest, charges $0.0375 per million input tokens and $0.15 for output — an entry-level rate. The previous generation stays in the catalogue at higher prices: Command R+ 04-2024 at $3.00 and $15.00, its August 2024 revision at $2.50 and $10.00. An older, pricier model still running in production isn't unusual for a vendor that guarantees API stability to its existing customers; it's still worth checking, before signing, which version the contract actually offers.
What the company promises, what the default doesn't cover
The security documentation highlights customer control over training and a SOC 2 Type II certification on the API infrastructure. That control does exist, but it's negotiated: an Enterprise account can request a DPA and exclude its data from any service-improvement use, while a trial or research account stays, by default, within a framework where inputs and outputs may feed internal research work. Processing region follows the same two-speed logic: the API infrastructure runs on Google Cloud in the United States, with transfers to Canada and the United Kingdom under the privacy policy, and it's only with a private deployment — customer VPC, on-premises, or Model Vault managed by Cohere — that a European region becomes possible.
Where it falls short
Two reservations to know before choosing Cohere for a sensitive project. The free trial key, first: it's explicitly banned in production, not a usable tier beyond testing. Second, without an Enterprise negotiation, no written guarantee fixes a processing region in Europe — a point that matters for a legal or healthcare sector required to document where its data lives.
Our recommendation
Cohere suits a technical team at a company ready to negotiate an Enterprise contract to get written guarantees on its data. Without that negotiating budget, or for a need for proven European hosting stated without detour, Scaleway or Hugging Face answer the question more directly.
Who it's for, and who it isn't
Suited for
- A development team at a company able to negotiate an Enterprise offer and a DPA to precisely govern its data use
- A developer building a document search or ranking feature (Embed, Rerank) rather than a simple conversational agent
Not a fit
- A structure that wants to test under real conditions before paying: the free trial key bans all commercial use, unlike a simple sign-up credit
- A small business that needs proof of European hosting by default, with no Enterprise negotiation or private deployment
Frequently asked questions
- Do I need to know how to code to use Cohere?
- Yes. Cohere sells an API for text generation, search (Embed), and ranking (Rerank): without a developer to integrate it, there's nothing to get out of it.
- Does my data train Cohere's models?
- Not by default for an Enterprise account, which can negotiate a DPA. For a trial or research account, the documentation warns that inputs and outputs may feed research work.
- Where is my data hosted with Cohere?
- The API infrastructure runs on Google Cloud in the United States, with possible transfers to Canada and the United Kingdom. Dedicated hosting in Europe requires an Enterprise offer or a private deployment (VPC, on-premises, or Model Vault managed by Cohere).
- How much does Cohere cost for a common model?
- Command R7B, the cheapest in the lineup, charges $0.0375 per million input tokens and $0.15 per million output tokens. Command R+ 08-2024, more capable, goes up to $2.50 and $10.
- Is Cohere a European alternative to OpenAI?
- No. Cohere is a Canadian company whose default infrastructure runs in the United States: it isn't a European sovereignty choice, despite its enterprise positioning.
Alternatives to Cohere
Platforms and APIs · AI agents
Google Vertex AI
Google cloud platform for deploying and orchestrating AI models (including Gemini) and agents in production.
Platforms and APIs · AI agents
Amazon Bedrock
AWS cloud service for accessing several AI foundation models through a single API and building agents.
Platforms and APIs · Language models
Hugging Face
The public repository of AI models: you can find, test and host almost everything that exists there.
Platforms and APIs
OpenRouter
Single gateway giving access to many AI models from different providers through one API, with automatic routing.
Our sources
- Pricing sheetaccessed on 30 July 2026
- Privacy policyaccessed on 30 July 2026
- Security pageaccessed on 30 July 2026
- Subprocessor listaccessed on 30 July 2026
- Terms of use (jurisdiction: Toronto)accessed on 30 July 2026
Prices and hosting terms change often. We recheck them periodically, but the publisher's page remains the only up-to-date reference.