Part of
Microsoft Foundry
Architecture

Microsoft Foundry Pricing: What Drives Your Monthly Bill?

Dark card: Microsoft Foundry pricing, source Microsoft Learn; a fine-tuned model, 720 hours of hosting in a 30-day month whether or not it is called, and before you update the budget, 1 full billing cycle to compare each meter with the estimate, highlighted in orange.

Azure AI Foundry, now named Microsoft Foundry, has no single price. Your monthly bill is the sum of the Azure services your build uses. That is why the Azure pricing calculator has no Foundry page. Microsoft's guide to planning and managing Foundry costs says Foundry "is composed of several optional Azure services". Two kinds of charge drive the bill, and Azure records each one as a meter, a line item on your bill. Usage meters bill what you consume. Examples are input and output tokens on a pay-as-you-go model deployment, a Code Interpreter session, or the vector storage behind file search. Tokens are the small chunks of text a model reads and writes. Capacity meters bill by the hour for as long as something is deployed, whether or not it handles a request. They cover provisioned throughput unit (PTU) hours on a provisioned deployment, a fine-tuned model's hosting, a compute instance and the GPUs of managed compute. On top of both come the other Azure resources your solution runs on, such as Azure AI Search, storage, networking and monitoring.

The model itself is billed per token or per hour, depending on the deployment type you choose. A pay-as-you-go deployment bills input and output tokens only when it is called. A provisioned deployment bills PTU hours on the capacity you deployed, with no separate per-token charge (Microsoft Learn, provisioned throughput billing). On a pay-as-you-go deployment, an idle week bills no tokens. On a provisioned one, an idle week bills the same PTU hours as a busy week.

Billed perAn idle weekBilling stops
Model tokensinput and output tokensbills no tokenswhen nobody calls
Provisioned deploymentPTU hours on the units you deployedthe same PTU hours as a busy weekwhen the deployment is deleted
Fine-tuned modelan hourly hosting fee on top of its input and output tokenswhether or not it is calledremove the deployments you no longer need
Compute instancea virtual machine that stays on until you stop itit keeps paying for its disk and for a load balancer billed per daydeleting the compute ends both

Tokens, the usage side of the bill

Language and vision models bill by the token. Microsoft sets the billing unit and rate by model, deployment type and meter. Read the rate for your own deployment on the Azure OpenAI pricing page (per the Foundry cost guide). Input tokens cover your prompt and system instructions. Output tokens cover the model's response. They are separate meters with their own rates. A standard deployment that nobody calls bills no tokens.

Two things make the input count larger than the prompt you wrote. In a chat, each request sends the conversation so far. Microsoft's guide to chat completion models says "With each question asked and answer received, the messages list grows in size." The prompt and the completion, the model's reply, have to fit the model's token limit. Keeping them inside it is the caller's job. You can count tokens and trim older turns, or let the Responses API manage the history. In retrieval, every passage you fetch to ground an answer goes into the prompt as input tokens. How much you retrieve is a cost decision as well as a quality one.

You pay your model's rate on every one of those tokens. The rate varies by model, so a task that a smaller model handles well costs less on that model. A cap on output tokens (the max_completion_tokens parameter) limits what each response can bill. For work that can wait, the Foundry cost guide names Flex processing. It gives Azure OpenAI deployments discounted token rates, with meters of their own. Size the token side of an agent before you build it. Multiply its runs a month by its model calls per run and its tokens per call. Then price the input and output tokens at your deployment's rates on the Azure OpenAI pricing page.

Capacity billed by the hour

A provisioned deployment reserves model capacity in provisioned throughput units. You pay PTU hours on the units you deployed, whether or not the deployment is handling requests (Provisioned throughput billing and cost management). It cannot be paused, and billing stops only when the deployment is deleted. For a steady production workload, Microsoft points you to an Azure Reservation. That is a one-month or one-year commitment to a number of PTUs at a discounted hourly rate. Microsoft says to create the deployments first and buy the reservation after, because a reservation does not guarantee capacity. Microsoft also advises against scaling hourly PTUs up and down with traffic in production. Its reasons are cost, and the risk that the capacity is gone when you want it back.

A fine-tuned model, one trained further on your own examples, adds charges that run whether or not it is called. Training is billed per token for supervised and preference fine-tuning, and by training time for reinforcement fine-tuning. Once deployed on Standard or Global Standard, the model pays an hourly hosting fee on top of its input and output tokens. Microsoft's monthly example for a fine-tuned chatbot counts 720 hours of hosting for a 30-day month (Manage fine-tuning costs in Microsoft Foundry). A Developer deployment has no hourly hosting fee, but it is meant for evaluation and is removed after 24 hours. Microsoft advises removing the deployments you no longer need, starting with those that carry hourly hosting, PTU or GPU charges.

Compute follows the same rule. Hub-based projects use a compute instance to run prompt flow. A compute instance is a virtual machine that stays on until you stop it. Idle shutdown, or a start and stop schedule, turns it off when nobody is working (Plan and manage costs for Microsoft Foundry hubs). Even stopped, it keeps paying for its disk and for a load balancer billed per day. Deleting the compute ends both. Models on managed compute bill GPU-hours while the deployment runs, even when no request arrives. Hosted agents are billed on the container compute they use per hour (Foundry Agent Service pricing).

Every hourly charge follows one rule from Microsoft's cost pages: the units deployed, times the hours deployed, times the rate per unit-hour. The units are PTUs for a provisioned deployment, GPUs for managed compute, or one for a fine-tuned model's hosting. A deployment left up for a 30-day month runs 720 hours, according to Microsoft's fine-tuning cost guide.

An hourly charge
What a deployment billed by the hour costs in a month
PTUs, GPUs, or 1 for a fine-tuned model's hosting
From the Azure pricing page
0 to 31
0 to 24
Hours deployed
720 hoursdays deployed × hours a day
Charge for the monthExample
$7,200.00Kept up all 720 hours of a 30-day month: $7,200.00
Example values; replace them with your own. The worked example: 10 units × $1.00 × 720 hours = $7,200.00. The rule is Microsoft Learn's, from the fine-tuning and provisioned throughput cost pages: units deployed × hours deployed × rate per unit-hour, billed whether or not the deployment handles a request.

Agent tools and the Azure resources around Foundry

Agents add usage meters of their own. Foundry Agent Service charges the input and output tokens of the base model behind each agent, so each agent you create bills its own model's tokens. Creating and running a prompt agent carries no charge of its own. Code Interpreter is charged per session, and a session stays active for one hour by default. Two threads calling it at the same moment open two sessions. File search is billed on the vector storage you keep, and agent data stays until you delete it (Foundry Agent Service FAQ). The pricing page bills that storage per GB per day, with the first GB free.

Then come the Azure resources the solution runs on. The Foundry cost guide says to watch these as well, since Foundry costs are "only a subset of your overall application or solution costs". Azure AI Search for retrieval, Blob Storage for project files, Key Vault, Container Registry and Application Insights each bill on their own pricing pages. On a hub-based project, the last four keep adding to the bill after the hub itself is deleted (per the hub cost guide). Bandwidth is charged by the data you move. Microsoft's Foundry Agent Service FAQ calls data transfer between peered virtual networks "costly", and does not recommend that setup for that reason. Sending logs to Azure Monitor and alerting on them adds charges too. Those charges show under those services and at subscription level. You will not see them if you scope cost analysis to the Foundry resource alone.

How to estimate the bill and hold it there

Microsoft's Foundry cost guide gives the order for getting from an estimate to a bill you can predict. Build an estimate in the Azure pricing calculator for every service in your architecture. Deploy a small test workload and send it representative traffic. In Cost Management, group costs by resource, then by meter. Compare each meter with the estimate and adjust your baseline budget. When this works, each assumption in your estimate maps to a meter you can see. Update the budget and its alert thresholds only after you have watched at least one full billing cycle.

Azure OpenAI has no hard spending limit, which OpenAI's own platform offers. Instead, a budget with alerts warns you when spend approaches or passes a threshold. An action group on the alert, a set of actions Azure runs when the alert fires, can start automation, which Microsoft notes takes custom development. After rollout, keep checking the cost by meter against the estimate. That is where a fine-tuned model's hosting hours show beside its tokens. In cost analysis, Azure OpenAI usage sits under the Cognitive Services service, so filter by the service tier Azure OpenAI. For models sold by Azure, project usage carries an automatic project tag, in preview. The tag splits a shared Foundry resource's bill by project.

When Ovidius builds on Microsoft Foundry, we estimate the running cost of your workflow during scoping. We check the usage estimate against the Azure bill during engineering. Before the first agent goes live, we build cost guardrails and approval rules into the foundation. If another team started your Foundry project, we begin with a technical assessment of what is built.

Frequently asked questions

What drives the cost of Microsoft Foundry?

Two kinds of charge drive it. Usage meters bill what you consume, such as input and output tokens on a pay-as-you-go deployment, Code Interpreter sessions and vector storage for file search. Capacity meters bill by the hour while something is deployed, whether or not it handles a request. They cover provisioned throughput units, a fine-tuned model's hosting, compute instances and managed compute GPUs. The other Azure resources of your solution, such as Azure AI Search, storage and bandwidth, add their own charges.

Is Microsoft Foundry free to use?

No. Foundry has no price of its own, so you pay for the Azure services your build uses. Creating and running a prompt agent carries no extra charge. The model tokens it uses are billed, and so is each tool, deployment and resource you add. Microsoft's Foundry cost guide gives this as the reason the Azure pricing calculator has no Foundry page. Foundry is made of several optional Azure services, each priced on its own.

Do input and output tokens cost the same in Azure OpenAI?

They are billed as separate meters, each with its own rate. Microsoft sets the billing unit and rate by model, deployment type and meter. Read both rates for your model and deployment type on the Azure OpenAI pricing page. In a chat, the input count grows with every turn, because each request sends the conversation so far. Microsoft's guide to chat completion models describes this.

Do provisioned throughput units cost money when they are not used?

Yes. A provisioned deployment is billed hourly on the PTUs you deployed, whether or not it handles requests. It cannot be paused, and billing stops only when you delete it. For a steady production workload, Microsoft's provisioned throughput billing guide recommends an Azure Reservation. That is a one-month or one-year commitment at a discounted hourly rate. Buy it after the deployments exist, because a reservation does not guarantee capacity.

Can you set a hard spending limit on Azure OpenAI?

No. Microsoft's Foundry cost guide says OpenAI offers hard limits and Azure OpenAI does not. Set budgets with alerts in Cost Management instead, for the subscription and the resource groups, and filter them to the resources you want to watch. An action group on a budget alert can start automation that acts on the overspend, which Microsoft notes takes custom development.

Why does a fine-tuned model cost money when nobody calls it?

It pays an hourly hosting fee for as long as it is deployed, on top of the tokens of each call. That applies on a Standard or Global Standard deployment. Microsoft's fine-tuning cost guide gives an example that counts 720 hours of hosting for a 30-day month. Remove the deployments you no longer need. You can also evaluate on a Developer deployment, which the same guide says has no hourly hosting fee and is removed after 24 hours.

Planning a Foundry build, or reading a bill that grew? Book a discovery call. A senior partner runs the call and tells you whether Foundry is the right platform for the job before anything is scoped.

More on this topic

Put your first workflow on the board

Answer a few questions about your team and what you want automated, then a senior partner scopes it with you on the call.