Cohere is a Toronto-based enterprise AI company whose commercial terms are issued by Cohere Inc.[4] It sells large language models (the Command family), retrieval models (Embed and Rerank), a document parser (Parse), a speech model (Transcribe) and two applications: North, an agentic AI platform, and Compass, a search and discovery system.[1] For a software asset manager the important point is that there is no single Cohere licence. The same model can be consumed under a click-through SaaS agreement, a Client License Agreement signed through a reseller, a marketplace end-user agreement on a hyperscaler, or an open-weight licence, and the terms, metrics and audit rights differ in each.[3]
Routes to buy and run Cohere
Cohere describes four deployment options: the Cohere platform, cloud AI services, private deployments in a customer cloud (VPC), and on-premises deployments, including air-gapped environments.[2] The platform is Cohere-hosted and reached through an API; cloud AI services are Cohere models hosted by providers such as Amazon Bedrock, Amazon SageMaker, Azure AI Foundry and OCI Generative AI.[2]
The legal centre lists the agreements that map to these routes: a Commercial SaaS Agreement, a Client License Agreement, an Azure AI EULA for Cohere models, channel partner marketplace terms, website terms of use and a service level objective.[3] The privacy and data documents are an Enterprise Data Commitments statement, a Privacy Policy and a Model Training Privacy Notice.[3]
Editions and offerings
API access. Every account receives a free, rate-limited Trial key, and a Production key is billed pay-as-you-go after an organization Owner completes the Go to Production workflow.[1][12] Generative models are priced per token, Rerank per search and Embed per token embedded; the detail is in Cohere API pricing, keys and rate limits. The catalog lists SKUs for the Trial key, the Production key, Command A, Embed 5 Pro and others.
Model Vault. Model Vault is a Cohere-managed, single-tenant inference environment in two types, Standard and Encrypted, billed per instance by model and performance tier.[10][1] The pricing page offers fixed or flex plans and hourly billing or monthly and annual commitments.[1]
North and Compass. North and Compass carry no list price; the pricing page tells buyers to get in touch for custom enterprise pricing.[1] The Client License Agreement shows how North entitlements are measured: by North Users or by tokens per minute.[5]
Private deployments. Cohere provides container images through a registry authenticated with a licence it issues through sales.[14]
Metrics
The metrics a licence manager will meet are listed here with their catalog rows.
| Metric | Where it applies | Counting basis |
|---|---|---|
| Chat tokens | Command models on the API | Billed input and output tokens, priced separately[1] |
| Rerank search unit | Rerank | One query with up to 100 documents[1] |
| Model Vault instance | Model Vault | Instance per model and tier, billed hourly or by commitment[1] |
| North User | North | A user designated to access North, active or not[5] |
| North tokens per minute (TPM) | North | Maximum tokens per minute reached during the term[5] |
| Cohere product instance | Client License Agreement | Instance reported in the Usage Report[5] |
| Permitted User | SaaS Agreement | Employee or contractor authorised by the customer[4] |
Counting and floors
There is no seat minimum on the pay-as-you-go API. The pricing page states that bills are issued at the end of every calendar month or when outstanding balances reach USD 250.[1] Trial keys, and production keys on newer Chat variants, are limited to 1,000 API calls a month, and each model and endpoint has its own requests-per-minute limit.[11] Under the Client License Agreement the North User count includes every designated user whether or not active, and a removed user leaves the count only 90 days after the account is decommissioned.[5] A vault is limited to three per organization unless Cohere raises it, with one to 25 replicas per model.[10]
Virtualization and deployment placement
Cohere does not publish virtualisation or partitioning rules. What it does publish is where the software may run. Under the Client License Agreement the client is responsible for its own environment, meaning an on-premises or private cloud environment owned or controlled by the client, and must remove Cohere products from it within five days of termination.[5] North may only be hosted in an environment meeting Cohere’s minimum requirements.[5] Cohere states that in private and third-party cloud deployments it does not receive customer prompts or generations.[8]
Cloud and marketplace use
Models bought through a hyperscaler marketplace fall under Cohere’s marketplace terms. The Azure AI EULA says fees and taxes are paid to the channel partner, that Cohere receives usage data from the partner and does not train on customer data, and that only Development Users may reach the models directly.[6] The models page lists Amazon SageMaker, Amazon Bedrock, Microsoft Azure and Oracle GenAI Service alongside Cohere’s own platform.[9]
Programs and policies
- Go to Production. A production key requires acknowledging the SaaS agreement and terms and declaring sensitive use cases, which may keep trial limits until approved.[12]
- Enterprise Data Commitments. Opt-out from training, 30-day log retention and zero data retention for approved enterprise customers.[8]
- Usage Policy. An acceptable-use policy referenced by the SaaS Agreement and the Client License Agreement.[13]
- Website Terms of Use. The page states it was last updated on 7 September 2022 and says the solution is designed for business use.[7]
Two things stand out for governance. The SaaS Agreement lets Cohere change its fees on seven days notice and bills at the then-current price list,[4] so price lists should be captured with a retrieval date. And the Client License Agreement carries a usage-report obligation and a ten-day audit right that survives the term for a year, which is relevant to any private deployment.[5]
Coverage in this wiki
- Cohere API pricing, keys and rate limits: per-token and per-search pricing, trial and production keys, rate limits and model lifecycle.
- Cohere Model Vault and private deployments: instance pricing, Fixed and Flex plans, Encrypted Vault, cloud services and private deployment licences.
- Cohere model licences and open weights: Apache 2.0 and CC BY-NC releases and the acceptable-use policies.
- Cohere agreements and data terms: the SaaS Agreement, Terms of Use, Client License Agreement, marketplace terms and data commitments.
- Advance Local Media v. Cohere: a pending United States copyright and trademark suit by news publishers.
Out of scope
Individually negotiated Order Forms, custom model and fine-tuning engagements, North and Compass prices, and the Service Level Objective, which the legal centre lists but whose page did not load its content in the retrieved copy.[3] Open-weight files hosted on third-party sites are governed by the licence shipped with the weights, which this article does not reproduce.