Cohere Model Vault is the vendor’s dedicated, managed route for running its models in production. Cohere describes it as an inference environment in which customers deploy and serve Cohere models in an isolated, single-tenant setup, with dedicated infrastructure and no operation of the serving stack by the customer.[1] It sits beside two other dedicated routes: Cohere models hosted on cloud AI services, and private deployments that a customer runs in its own cloud or data centre.[5] For software asset management the three routes have different commercial units, different contracts and different answers to the question of who holds the data.
Editions: Standard and Encrypted vaults
Model Vault comes in two types. A Standard Vault is a Cohere-managed, single-tenant deployment with data protected in transit and at rest. An Encrypted Vault adds confidential computing, so that prompts and responses stay protected inside hardware-backed trusted execution environments, with verifiable remote attestation.[1] Both are used standalone over the API or as the inference backend for North, and both are described as supporting GDPR, HIPAA and SOC 2 compliance, with attestation evidence for the Encrypted type.[1] The type is chosen when the vault is created and cannot be changed afterwards.[4]
Encrypted Vault is currently in beta and is offered free of charge to design partners. Cohere says full pricing details will be shared as the product matures.[3] A buyer evaluating it should record the beta status in the asset register, because a free design-partner arrangement is not a commitment on future price. The catalog rows are Model Vault Standard and Model Vault Encrypted.
Supported models
A Standard Vault supports the Command family (including reasoning, multimodal and translation variants), North generative models including North Mini Code, Embed, Rerank, Parse, a Compass bundle of Embed, Rerank and Parse, Cohere Transcribe and customized Cohere models.[11] Embed and Rerank are generally available self-serve; generative models, bundles, speech and customized models may require a waitlist.[11] Cohere’s model cards for Command A+ and North Mini Code state that production use is through Model Vault.[10]
Metrics: the instance and the vault
The billing metric is the instance. Pricing is per instance and depends on the model chosen and its performance tier, with tiers shown as size letters S, M, L and XL.[2][1] Within a vault the customer sets a minimum and maximum replica range of 1 to 25 per model, with autoscaling available.[1] An organization is limited to three vaults unless a Cohere representative raises the limit.[4] The catalog rows are Model Vault instance and Model Vault.
Because a vault is single-tenant, Cohere states that there are no rate limits on it and that the customer pays only for what it uses.[1] That second statement is qualified by the plan types below: under a Fixed plan the instances are a commitment, not a metered amount.
Counting and floors: Fixed and Flex plans
Cohere’s Standard Vault pricing offers two pricing models, both on a monthly or annual commitment.[2]
| Feature | Fixed | Flex |
|---|---|---|
| Capacity | Fixed number of instances, no autoscaling | Minimum baseline instances plus autoscaling |
| Sizing | Set by a sizing exercise or production trial | Not applicable |
| Autoscaling | Not applicable | Scales on request rate and agreed latency SLOs |
| Pause or resume | Yes | Yes |
| Overages | Not applicable | Additional capacity billed per instance-hour |
| Maximum capacity | Not applicable | Maximum instance cap per model |
Under both plans a model can be paused and restarted to save on cost.[2] The main pricing page adds that billing can be calculated hourly or through longer-term commitments, monthly or annual.[7] The catalog row is Model Vault Fixed and Flex plans.
Standard Vault rates
All rates are per instance. The Standard Vault pricing page lists these rates for retrieval and speech models.[2]
| Model | Tier | Hourly | Monthly | Annual |
|---|---|---|---|---|
| Embed 3 and Embed 4 | Small | USD 4.00 | USD 2,500 | USD 25,000 |
| Embed 4 | Medium | USD 5.00 | USD 3,250 | USD 32,500 |
| Embed 5 Fast and Pro | Small | USD 3.00 | USD 2,000 | USD 20,000 |
| Embed 5 Fast and Pro | Medium | USD 5.00 | USD 3,250 | USD 32,500 |
| Rerank 3.5, Rerank 4 Fast, Rerank 4 Pro | Medium | USD 5.00 | USD 3,250 | USD 32,500 |
| Rerank 4 Pro | Large | USD 10.00 | USD 6,500 | USD 65,000 |
| Parse 5 | Medium | USD 4.00 | USD 2,500 | USD 25,000 |
| Parse 5 | Large | USD 8.00 | USD 5,000 | USD 50,000 |
| Parse 5 | XL | USD 7.00 | USD 4,300 | USD 43,000 |
| Cohere Transcribe | Medium | USD 3.75 | USD 2,500 | USD 25,000 |
| Cohere Transcribe | Large | USD 7.50 | USD 4,750 | USD 47,500 |
Generative models are priced per instance-hour only on that page: Command A, Command A Vision and Command A Translate are USD 40.00 on an L tier and USD 48.00 on XL, Command A Reasoning USD 48.00 and USD 57.50, Command A+ USD 17.50 and USD 32.50, and North Mini Code USD 7.50 and USD 10.50. Monthly and annual rates for these follow the same plans but are given on request.[2] The Cohere pricing page shows a shorter rate table for the same Embed, Rerank and Parse models, with Parse 5 offered in Medium and XL tiers.[7] Where the two pages differ, record which page was used and the date.
Cohere Transcribe’s model card repeats that pricing is calculated per hour-instance with discounted plans for longer commitments.[13]
Cloud AI services
Cohere states that it is cloud-agnostic and lists four cloud AI services: Amazon Bedrock, Amazon SageMaker, Azure AI Foundry and OCI Generative AI.[5] In these, Cohere’s models are hosted on the provider’s infrastructure.[5] The commercial relationship is with the provider. Under Cohere’s marketplace terms the customer pays fees and taxes to the channel partner, only Development Users may directly access the channel partner APIs, and Cohere receives usage data from the partner but not customer data.[12] The catalog rows are Cloud AI Services and Development User.
Private deployments
Private deployments let an enterprise run the Cohere stack on its own compute. In the cloud, models can be deployed in the customer’s AWS environment through its own VPC; on-premises deployments include air-gapped environments.[5] Cohere notes that VPC deployments give tighter control and compliance and a higher management burden than managed cloud services.[5]
Licence and image access
A private deployment requires a licence and an image tag provided by Cohere through sales.[6] Cohere provides container images through a registry authenticated with that licence, and customers may replicate the images inside their environment to avoid runtime network access from within the cluster.[6] The licence ID acts as the credential, so it should be recorded and protected like a licence key. The catalog row is Private Deployments.
Contract obligations in a client environment
When Cohere software is installed in a client environment, the Client License Agreement applies. It requires a usage report for each Cohere product instance, lets Cohere or its partner audit use on ten days written notice during the term and for one year after it, and requires the client to remove Cohere products from its environment within five days of termination and certify removal on request.[9] The metric is Cohere product instance. The agreement’s North Product Schedule adds that North may only be hosted in an environment that meets Cohere’s minimum requirements.[9]
Data in private and third-party deployments
Cohere states that in third-party cloud AI platforms and private deployments it does not receive customer prompts or generations, and that no Data Processing Addendum is required for those routes.[8] It may still receive usage data under its commercial agreements.[8]
Programs
Compass is a bundle in Model Vault and a separately priced application, and Cohere gives no list price for it or for North, directing buyers to custom enterprise pricing.[7] For North, the Client License Agreement counts North Users or tokens per minute; see Cohere agreements and data terms.
Out of scope
Hardware sizing for private deployments, GPU availability by region, Order Form pricing for private deployments, and the Service Level Objective for the SaaS service. Cohere says that supported GPUs vary by model, capacity and region and must be confirmed with Cohere.[11]