Billing concepts for Microsoft Foundry

Microsoft Foundry bills you for the underlying Azure services and models that you use. How each charge accrues depends on its billing model, such as pay-as-you-go or provisioned throughput. For models, it also depends on whether Azure or a partner sells the model.

This article explains each billing model, how Foundry Models are billed, and where the charges appear in Azure Cost Management. Use it to choose a billing model for a workload and to recognize each charge on your bill.

Important

Items marked preview in this article are currently in preview. This preview is provided without a service-level agreement, and Microsoft doesn't recommend it for production workloads. Certain features might not be supported or might have constrained capabilities. For more information, see Supplemental Terms of Use for Microsoft Azure Previews.

How Foundry charges accrue

When you create a Foundry resource, you pay for the Azure services that you use. Costs vary by service and feature. The services you pay for can include:

  • Azure OpenAI
  • Azure Speech in Foundry Tools
  • Azure Language in Foundry Tools
  • Azure Vision in Foundry Tools
  • Azure Document Intelligence in Foundry Tools
  • Azure Content Safety in Foundry Tools

Foundry resources run on Azure infrastructure. Some deployments accrue charges for as long as they exist, even when they aren't in use. Examples include provisioned throughput deployments and fine-tuned models on deployment types that have an hourly hosting fee.

Your Foundry costs are only part of your overall solution cost. Monitor costs for all the Azure resources that your application or solution uses. For more information, see Other costs that accrue.

For current rates, see Foundry Tools pricing and Microsoft Foundry pricing.

Billing models

A billing model determines how you pay for usage: as you go, through a commitment, or for dedicated capacity. The following table compares the billing models that Foundry supports.

Billing model Best for How it's billed Learn more
Pay-as-you-go Usage that you pay for as it happens, with no long-term commitment. By your usage of each Azure service. For models, usage is typically metered in tokens. Token-based pricing
Commitment tiers Predictable costs when you can commit to a level of usage. A fixed fee for a committed amount of usage. Usage above the commitment is billed at an overage rate. Commitment tier pricing
Provisioned throughput, with optional Azure Reservations Production workloads that need dedicated capacity and predictable latency. Hourly, for each deployed provisioned throughput unit (PTU), whether or not the deployment handles requests. Azure Reservations provide a discounted rate for sustained workloads. Provisioned throughput billing
Batch Large-scale, high-volume jobs that don't need immediate responses. At a lower cost than the Global Standard deployment type, with a 24-hour target turnaround. If you cancel a job, you pay for any work that completed. Batch processing
Flex processing (preview) Noninteractive, lower-priority work that tolerates slower responses and occasional resource unavailability, such as model evaluations and data enrichment. At discounted token rates, with dedicated meters for cost analysis. You choose Flex processing for each request. Flex processing
Priority processing Time-sensitive workloads that need consistent low latency, such as business-hour or bursty traffic. Pay-as-you-go at priority processing rates, with no long-term commitment. You enable it on a deployment or for each request. Priority processing

A commitment tier provides a usage quota. If you exceed that quota, you pay for the extra usage at the overage amount. The Azure portal shows the overage amount when you buy the commitment tier.

Provisioned deployments can't be paused. Billing stops only when you delete the deployment. For more about dedicated capacity, see What is provisioned throughput for Foundry Models?

For current model rates, see the Azure OpenAI pricing page.

How Foundry Models are billed

How you're billed for a Foundry Model depends on who sells it. The offer determines where charges appear in Azure Cost Management and whether Azure Prepayment can pay for them.

Model offer Billed through Where charges appear in Cost Management Azure Prepayment applies
Foundry Models sold by Azure, including Azure OpenAI Microsoft, directly Typically as model-related meters associated with your deployed resources. Yes
Foundry Models from partners and community Azure Marketplace The cost meters display under the resource group instead of the Foundry resource. Some providers' models are displayed as meters under Global resources. No
Claude models Azure Marketplace, in Claude Consumption Units (CCU) Under the Foundry resource, with model and usage-type breakdowns. No

You can use Azure Prepayment (previously called monetary commitment) credit to pay for charges for Models sold by Azure. You can't use it for other providers' models, because they're billed through Azure Marketplace.

CCU spend for Claude counts toward your Microsoft Azure Consumption Commitment (MACC). Claude deployments created before CCU billing became generally available continue to bill on their existing per-model plan. For details, see Claude consumption units (CCU) billing.

To find these charges in cost analysis, see Analyze Foundry costs in Azure Cost Management.

How agents are billed

When an agent in Foundry Agent Service runs, you pay for the model calls it makes and for the tools and resources it uses. Each item bills on its own meters, so an agent's total cost spans several meters in Azure Cost Management.

What's billed How it's billed Included in agent costs in the Foundry portal
Model calls Tokens that the agent's model deployment processes, on that deployment's meters and billing model. Tool calls add model turns, which add tokens. Yes, for standard deployments. For provisioned throughput, see Understand the cost data.
Model calls through model router Model router charges for input tokens at its own rate on the pricing page. In addition, each request is billed at the rates of the underlying model that model router selects for it. For details, see Model router billing. Yes
Tools Some tools have charges beyond model token fees, for example Code Interpreter and file search. Server-side tools, such as toolboxes and MCP servers, bill through their own resources. Not during public preview. Model tokens from tool calls count as model calls.
Hosted agent runtime Based on the CPU and memory that hosted agents consume during active sessions. For details, see Hosted agents pricing. Not during public preview
Dependent resources, such as Foundry IQ (Azure AI Search), storage, and Application Insights Each resource bills on its own meters, under its own service. Not during public preview

During public preview, agent costs in the Foundry portal cover model calls only. To see the billed cost of every part of an agent solution, use Azure Cost Management. For current rates, see Foundry Agent Service pricing.

Token-based pricing

Language and vision models process inputs by breaking them into tokens. Text, image, and audio workloads can all use token-based metering.

Billing typically uses separate meters for input and output tokens, and some models also have a meter for cached tokens. The billing unit and rate can vary by model, deployment type, and meter. Check the pricing page for the exact meter names and units for your deployment.

For current rates, see the Azure OpenAI pricing page.

Fine-tuned models

Azure OpenAI fine-tuned models incur charges for training, for hosting or provisioned capacity while deployed, and for inference.

Charge When it accrues How it's billed
Training When you train the model. Per token or per hour, depending on the model.
Hosting While the model is deployed on a deployment type that has an hourly hosting fee, such as Standard or Global Standard, even if you don't use it. Developer deployments don't have an hourly hosting fee. Per hour.
Provisioned capacity While the model is deployed with provisioned throughput, regardless of usage. Hourly, for each allocated PTU.
Inference When you call a deployment that's billed per token. Per input and output token. Provisioned throughput deployments don't have a separate per-token charge.

Because hourly hosting and provisioned capacity charges accrue for as long as a fine-tuned model is deployed, monitor hosted fine-tuned model costs closely. Remove or scale down deployments that you don't need.

For current fine-tuning rates, see Azure OpenAI pricing. For other hosting options and their costs, see Manage fine-tuning costs in Microsoft Foundry. For the deployment lifecycle and cleanup policies, see Fine-tune a model with supervised fine-tuning.

Requests that fail

A failed request is billed only if it includes billable processing, and the billing depends on the specific meter's behavior. The HTTP status code alone doesn't determine whether you're charged.

  • On provisioned throughput deployments, you're billed for the deployed PTU count regardless of utilization, so a failed request doesn't change the charge.
  • Requests that a cost control blocks are rejected before processing, so they aren't billed.

To reconcile billed usage, use the meter data in Azure Cost Management and service metrics. Treat your invoice and meter records as the source of truth.

Other costs that accrue

Enabling capabilities such as sending data to Azure Monitor Logs and alerting incurs extra costs for those services. These costs appear under those services and at the subscription level. They don't appear when you scope cost analysis to your Foundry resource.

Dependent resources in your solution also accrue their own charges. For example:

  • Application Insights, which you connect to a project to see usage against cost controls.
  • Foundry IQ, if your solution uses search.
  • Storage and networking resources.

To see your full solution cost, include these resources in your Cost Management scope.

Billing by offer

Some offers have their own billing articles: