Cohere
Cohere is a Canadian provider of enterprise foundation models: Command for generation, Embed and Rerank for retrieval, Transcribe for speech. Its North and Compass platforms and private VPC, on-premises or Model Vault deployments keep data under customer control.
What is Cohere?
Cohere Inc. is a Canadian company founded in 2019 in Toronto that builds foundation models and AI platforms for organisations. The homepage states the positioning bluntly: "Own your AI. Your data. Your infrastructure. Cohere keeps it that way." Three pillars structure the offer, namely security, deployment and customization.
On the model side, Command is the generative family, presented as delivering powerful agentic performance with minimal compute overhead and bringing unified reasoning, tool orchestration and multimodal intelligence together in a single model, with 49 languages announced on the homepage and 48 on the Models Overview page. Embed handles semantic representation across 23 languages with a 128K context window and multimodal input, image processing being priced separately. Rerank gives search quality a semantic boost, where one search unit is a single query covering up to 100 documents, within a 32,768 context window. Transcribe does speech recognition in 14 languages, processes audio in 30-second blocks, is built to hold up in real conversational environments and plugs into generative and retrieval systems. North Mini Code is an agentic coding model aimed at practical software engineering.
Two platforms sit on top of the models. North, launched in 2025, is a productivity AI workspace with a unified interface, intelligent search and agents that handle both routine tasks and complex workflows; an instant demo is available. Compass covers enterprise search and discovery with prebuilt data connectors, document parsing and a managed index. Model Vault is a dedicated, logically isolated inference platform fully managed by Cohere, and the models also run on any AI or ML cloud: Oracle Cloud Infrastructure, Azure, AWS and Google Cloud.
The compliance base is documented: SOC 2 Type II with an annual audit, ISO 27001, ISO 42001 for AI management systems and UK Cyber Essentials. Seven sector Solutions pages address technology, financial services, healthcare and life sciences, manufacturing, energy and utilities, public sector and telecommunications, while named partners include Fujitsu for Japanese models, LG CNS in South Korea, Dell for on-premises deployments and AMD on hardware.
Research runs alongside the commercial range through Cohere Labs, an open research lab started in 2022 with more than 4,500 community members, over 100 published papers, the Aya models covering more than 100 languages, Global MMLU and The Leaderboard Illusion. Offices are in Toronto, New York, London, San Francisco, Montreal, Paris and Seoul.
What it does
- Generate text and reason over it with the Command models, which combine agentic behaviour, tool orchestration and multimodal input
- Vectorise and search documents with Embed 4, multimodal and with a 128K context window
- Re-rank search results by relevance with Rerank 4, available in Fast and Pro variants
- Transcribe audio into text with Transcribe, covering 14 languages in 30-second blocks
- Automate tasks and workflows with the North agents and North Automations
- Search enterprise data with Compass, using prebuilt connectors, document parsing and a managed index
- Generate code with North Mini Code, then customise any model on proprietary data and deploy it in a VPC, on-premises or in Model Vault without exposing that data
When to use Cohere / When not to
A quick filter to help you decide if Cohere is the right fit.
When to use Cohere
- Large regulated enterprises that need control over data residency and over where their models actually run
- Platform and AI engineering teams deploying language models inside their own VPC or on-premises infrastructure
- Developers and machine learning engineers consuming Command, Embed and Rerank directly through the API
- Public sector bodies and regulated industries covered by the seven sector Solutions pages: financial services, healthcare and life sciences, manufacturing, energy and utilities, telecommunications and technology
- Academic and open science research teams working with Cohere Labs, the Aya models, the grant programmes and the Scholars Program
When not to use Cohere
- Consumers and individual users: there is no mobile app and no consumer plan, and the terms of use state that Cohere Products are not intended to be used by minors (children) and are not for personal or household use
- Organisations that require EU data residency on a hosted service, since the SaaS is hosted only on Google Cloud Platform servers in US-Central
- Teams planning to run production traffic on the free trial key, which is rate limited and not permitted to be used for production or commercial purposes
- Healthcare organisations expecting a HIPAA business associate agreement that covers the hosted SaaS products: the BAA is restricted to custom model development
- Buyers who need public, self-serve pricing for North or Compass, as both are only quoted through enterprise sales
How to use Cohere
A typical end-to-end flow, from setup to results.
- Create an account from the Get an API key link on the homepage, which leads to the Cohere dashboard registration page
- Collect the Trial API key, created automatically with the account and visible in the API Keys section of the dashboard
- Test the models in the Playground and follow the developer documentation on docs.cohere.com
- Build up skills with LLM University and the Cookbooks
- Pick a model according to the performance and latency trade-off, as the pricing FAQ recommends
- Join an organisation if billing has to be shared, since a personal account only gets trial keys
- Move to production with Owner privileges: open the Billing and Usage page, use the Get Your Production key button and complete the Go to Production workflow
- For North or Compass, request a demo: an instant demo is available for North, otherwise go through contact sales
- Choose a deployment mode between SaaS, Model Vault, public or hybrid cloud, and private (VPC or on-premises)
- Configure Model Vault from the Cohere dashboard, which the deployment comparison table lists as available instantly
Pros & Cons
Pros
- Data sovereignty: models can run in the customer's own VPC or on-premises, with Cohere holding no access to the infrastructure or the data
- Explicit opt-out from model training at any time, stated on the Security page
- Documented compliance base: annual SOC 2 Type II, ISO 27001, ISO 42001, UK Cyber Essentials, a DPA with Standard Contractual Clauses and a post-Schrems II Transfer Impact Assessment
- Subprocessor list published and consultable in the Trust Center
- Per-token model prices published, with a very low entry point at USD 0.0375 per 1M input tokens for Command R7B
- Free trial API key created automatically, with no payment method required
- One vendor covering generation, retrieval, reranking, transcription and code, with open weights for Command A+ and North Mini Code, both listed with Model download at 0
Cons
- SaaS hosting in the United States only: Google Cloud Platform servers in US-Central, with no servers used outside the US
- No public pricing for North or Compass, so both require going through sales
- Model Vault has a high entry point, from USD 3.75 per hour per instance, with a published table running from USD 4.00 per hour (USD 2,500 per month) to USD 10.00 per hour (USD 6,500 per month)
- The trial key is rate limited and not permitted to be used for production or commercial purposes
- The HIPAA business associate agreement is restricted to custom model development and excludes the hosted SaaS products
- The DPA and the SOC 2 report are only released after a non-disclosure agreement has been signed
- No mobile app, and a product aimed at organisations that excludes personal or household use; the site's terms of use are dated 7 September 2022, well before the 1 May 2026 privacy policy
Pricing & Plans
A free entry point exists: the Trial API key is created automatically with the account, requires no payment method, and is both rate limited and not permitted for production or commercial purposes. Production keys are charged on a pay-as-you-go basis, invoiced at the end of the calendar month or as soon as USD 250 of usage has accrued. The lowest published paid price point is Command R7B, at USD 0.0375 per 1M input tokens and USD 0.15 per 1M output tokens. Dedicated inference through Model Vault starts at USD 3.75 per hour per instance. North and Compass are subject to enterprise pricing on quote. All published amounts are in USD.
- Trial API key — free
- rate limited
- non-commercial use only
- Production API key — pay-as-you-go per token
- billed at the end of the calendar month or once USD 250 of usage has accrued
- Command R — USD 0.15 per 1M input tokens
- USD 0.60 per 1M output tokens
- 128K context
- Command R7B — USD 0.0375 per 1M input tokens
- USD 0.15 per 1M output tokens
- 128K context
- Command A+ — API key 0
- Model download 0
- Embed 4 — USD 0.12 per 1M (cost)
- USD 0.47 per 1M (image cost)
- 128K context
- 23 languages
- Rerank 4 Fast — USD 2.00 per 1K searches
- 32
- 768 context
- Rerank 4 Pro — USD 2.50 per 1K searches
- 32
- 768 context
- Transcribe — 14 languages
- 30-second blocks
- North Mini Code — free
- with downloadable weights
- Model Vault — from USD 3.75 per hour per instance
- Embed 4 Small USD 4.00 per hour or USD 2
- 500 per month
- Embed 4 Medium
- Rerank 3.5 Medium
- Rerank 4 Fast Medium and Rerank 4 Pro Medium USD 5.00 per hour or USD 3
- 250 per month
- Rerank 4 Pro Large USD 10.00 per hour or USD 6
- North — enterprise pricing on quote
- Compass — enterprise pricing on quote
- Legacy models still priced for existing customers — Command USD 1.00 / 2.00
- Command-light USD 0.30 / 0.60
- Command R 03-2024 USD 0.50 / 1.50
- Command R+ 04-2024 USD 3.00 / 15.00
- Command R+ 08-2024 USD 2.50 / 10.00 per 1M input / output tokens
- Aya Expanse 8B and Aya Expanse 32B through the API — USD 0.50 per 1M input tokens
- USD 1.50 per 1M output tokens
Data, GDPR & hosting
A consolidated view of how Cohere handles your data.
GDPR overview
The Trust Center lists the GDPR, CCPA and HIPAA among its compliance frameworks and states that Cohere, operating in Canada, the EU and the US, designed its services with privacy by design in mind and has processes to help customers meet their own obligations. A multi-jurisdictional DPA is prepared, incorporating the Standard Contractual Clauses approved by the European Commission on 4 June 2021, and a Transfer Impact Assessment was carried out with EU counsel after Schrems II. GDPR compliance training is delivered through Secureframe. The privacy policy, last updated on 1 May 2026, sets out the legal bases under the GDPR and UK GDPR, adds a section for European residents, relies on adequacy decisions for UK and EEA transfers to Canada, and commits to notifying public authority access requests unless legally prohibited. A Privacy Officer is reachable at privacy@cohere.com. No Article 27 EU representative is named.
Who owns the data?
The publisher is Cohere Inc., a Canadian company subject to Canadian federal privacy laws. Section 9 of the terms of use keeps all right, title and interest in the Cohere Solution, its source code and the materials provided as the sole property of Cohere or its third-party suppliers; section 10 grants users only a non-exclusive, non-transferable, non-sublicensable and revocable licence. Enterprise customer data, meaning inputs and outputs, is excluded from the privacy policy and governed instead by the enterprise agreement and, where signed, the DPA. Feedback ceases to be confidential, Cohere obtaining a worldwide, royalty-free, perpetual and irrevocable licence over it. The privacy policy dates from 1 May 2026 and the terms of use from 7 September 2022, both governed by Ontario law.
Reuse rights
The licence granted to the end user is non-exclusive, non-transferable, non-sublicensable and revocable, and section 9 of the terms of use reserves every right that is not expressly granted. The Cohere name and logo are trademarks of Cohere Inc. and may not be used without express written consent. Feedback sent to Cohere is covered by a perpetual and irrevocable licence, with no compensation. Enterprise users keep control over whether their data is used for training and may request a DPA, whereas on trial and research accounts Cohere collects the inputs submitted and the outputs generated and may use them. Two models carry open weights: Command A+ and North Mini Code are listed on the pricing page with API key and Model download both at 0, so their weights can be downloaded. No clause on the ownership of model outputs was found in the terms of use published on the site.
Data retention & training
Hosting summary
Cohere Inc. is a Canadian company, and personal information transferred from the UK or the EEA to Cohere in Canada is made pursuant to adequacy regulations. Hosting itself is American: the Trust Center states that the hosting centres are on Google Cloud Platform servers located in US-Central and that no servers outside the US are used. In certain circumstances the offering can be configured so that no customer data is stored on Cohere's systems, the data then being ephemeral, transiting briefly and purged immediately after processing at the cost of some functionality. Transfers from the EU to the US rely on the Standard Contractual Clauses of 4 June 2021 and on a post-Schrems II Transfer Impact Assessment, whose measures include encryption at rest and in transit, administrative access control, logging and alerting. The nine published subprocessors are all located in the USA: Google Cloud for infrastructure, plus Fullstory, LaunchDarkly, New Relic, Retool, Sentry IO, Segment, Sendgrid and Vercel. In private deployments and on third-party clouds, data stays in the customer's or the cloud provider's environment and is never accessible to Cohere. The resolved IP address 76.76.21.21 is an anycast CDN node, not a storage location.
Things to keep in mind
Risks and trade-offs to weigh before adopting Cohere.
- Data residency: the hosted service runs in the United States only, which can clash with an EU residency requirement, and EU to US transfers remain transfers outside the EU even when framed by Standard Contractual Clauses and a Transfer Impact Assessment
- The most useful compliance documents, the DPA and the SOC 2 report, require a non-disclosure agreement before they can be read, so due diligence cannot be completed from the public site alone
- Training opt-out is a setting rather than a default: misconfigured, your data can end up feeding model improvement
- The ephemeral configuration that stores nothing on Cohere's systems comes at a price the site warns about, namely reduced functionality and reduced support
- Model outputs need checking, since no accuracy is guaranteed and the terms of use exclude liability broadly
- Single-vendor dependency across generation, retrieval and transcription, with the legacy models already on their way out
- The site's terms of use have not been revised since September 2022, while the product range has changed considerably
Setup & Integrations
Technical difficulty
API access is immediate: registering creates a trial key automatically and the Playground opens at once, so first calls need only basic HTTP skills. Going to production adds Owner privileges and the Go to Production workflow. The deployment comparison table shows SaaS and Model Vault as available instantly, private and public or hybrid cloud in under a day. A private VPC or on-premises deployment is the demanding path and assumes internal infrastructure and skills. North and Compass ship as an out-of-the-box experience with prebuilt connectors. Documentation, LLM University, Cookbooks and Discord support the learning curve.
Deployment
Integrations
Supported languages
Behind Cohere
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
How do I get a Cohere API key?
What is the difference between a trial key and a production key?
How do I obtain a production key?
When am I billed for my usage?
Where are Cohere's servers hosted?
Does Cohere train its models on my data?
Is a Data Processing Agreement available?
Which security certifications does Cohere hold?
Will Cohere sign a HIPAA business associate agreement?
Which deployment modes are available?
Should you pick Cohere?
Cohere is an enterprise-first foundation model provider whose whole argument is sovereignty and security rather than consumer reach. A single vendor covers generation with Command, embeddings with Embed, reranking with Rerank, speech with Transcribe and code with North Mini Code, then adds two platforms on top: North for agentic productivity and Compass for enterprise search and discovery.
Its main strength is letting the buyer choose where inference happens, between a private VPC, on-premises, the Cohere-managed Model Vault, a public or hybrid cloud and the hosted SaaS, combined with an explicit opt-out from model training. The compliance base is solid and documented, with annual SOC 2 Type II, ISO 27001, ISO 42001, UK Cyber Essentials, a DPA carrying the Standard Contractual Clauses and a Transfer Impact Assessment conducted after Schrems II. Two caveats temper that record: the DPA and the SOC 2 report are only released under NDA, and the hosted SaaS runs exclusively on Google Cloud servers in US-Central, a real limit for any European organisation requiring EU data residency without going private.
Pricing is transparent where developers care most, with published and low per-token API rates, while Model Vault instances start at USD 3.75 per hour and North and Compass are quoted by sales only.
Behind the products sits a genuine open research effort through Cohere Labs, the Aya models and more than 100 published papers, plus industrial partnerships with Fujitsu, LG CNS, Dell and AMD. Founded in 2019 in Toronto and backed by nearly a billion dollars raised over four rounds, Cohere is a credible option for regulated enterprises, platform teams and public sector buyers who need to control where their data and their models live. It is not a fit for consumers, for personal use, or for anyone who needs EU-hosted SaaS.
- Choosing a selection results in a full page refresh.
- Opens in a new window.