12.3 Foundry Tools Subscription Models

Microsoft’s adoption skill for this topic is Understand Foundry Tools subscription models, including pay-as-you-go and commitment tiers. Product names on the July 22, 2026 blueprint are Microsoft Foundry and Foundry Tools (older docs may still say Azure AI Foundry or Cognitive Services). You are not asked to recite SKU dollars. You are asked to pick the billing model, honor the multi-service restriction, and explain how a leader watches cost in Azure Cost Management and the Foundry portal.

An operations VP who approved “one Foundry resource for everything” often learns the restriction the expensive way: the team wanted a Speech commitment for a call-center transcript workload, but the resource was a multi-service Foundry Tools or Microsoft Foundry account. Microsoft’s commitment-tier article states the rule in one sentence: You cannot use commitment tier pricing with a Foundry Tools or a Microsoft Foundry multi-service resource. You must use resources dedicated to single services, such as a Speech or Translator resource.

Pay-as-you-go: billed per usage of each service

Microsoft’s Plan and manage costs article for Foundry lists common billing approaches. Pay-as-you-go (Serverless API) means you are billed according to usage of each Azure service. Foundry is not a single calculator line: Microsoft says Foundry does not have a dedicated page in the Azure pricing calculator because it is composed of optional Azure services. Finance estimates by adding products such as Azure Speech in Foundry, Azure Language in Foundry, Azure AI Search, Azure Vision in Foundry, Document Intelligence, Content Safety, and Azure OpenAI / Foundry Models, then validating meters after a small test workload.

When a Foundry resource exists, you pay for the services you use. Costs vary by service and feature. Language and vision models typically bill in tokens (text, and for some workloads image or audio). Rates vary by model, deployment type, and meter. Microsoft tells you to read the pricing page for meter names rather than inventing units. HTTP status codes alone do not decide whether a call is billed; billed processing and meter behavior do. Reconcile against Cost Management and the invoice.

Foundry Models sold by Azure (including Azure OpenAI) are billed by Microsoft and show up as model-related meters. Fine-tuned Azure OpenAI models are charged three ways:

1. Training — per token or per hour, depending on the model.
2. Hosting — hourly cost per deployed model, even if unused.
3. Inference — per 1,000 tokens of input and output when the model is called.

That hosting line is the classic “we turned it off in the app but left the deployment” surprise. Microsoft tells teams to remove or scale down idle fine-tuned deployments. Partner or community models served through Azure Marketplace appear under the resource group, not always under the Foundry resource itself, and Azure Prepayment (formerly monetary commitment) can pay for Azure-sold models but cannot pay Marketplace provider-model charges.

Project-level chargeback is in preview: every Foundry project is tagged with a project tag on usage. Filter Cost Analysis by that tag. Microsoft currently documents this for Models sold by Azure only, not for Marketplace models.

Other costs accrue outside the Foundry resource scope: Azure Monitor Logs, alerting, storage, and networking. If finance scopes Cost Analysis only to the Foundry account, those lines disappear from the screenshot and reappear on the subscription invoice. Azure OpenAI usage in Cost Management often sits under the broader Cognitive Services classification; Microsoft documents a Service tier: Azure OpenAI filter when you need that slice.

Microsoft is also explicit that Azure OpenAI does not currently provide OpenAI’s hard budget limit that stops spend. Budgets send notifications and can start action groups; they are not a complete off switch without extra automation.

Commitment tiers: fixed fee, eligible features, overage

Commitment tiers let you commit to using listed Foundry Tools features for a fixed fee, which Microsoft describes as discounted versus Standard (PAYG) pricing and as a way to make workload cost predictable. Eligible features currently documented on the commitment-tier article include:

• Speech to text (Standard)
• Text to speech (Neural)
• Text Translation (Standard)
• Language Understanding standard (Text Requests)
• Azure Language: Sentiment Analysis, Key Phrase Extraction, Language Detection, Named Entity Recognition (NER)
• Vision in Foundry Tools — OCR
• Document Intelligence — Custom/Invoice

Microsoft points to the Foundry Tools pricing page for rates. This chapter does not invent SKU prices; overage is “the overage amount mentioned in the commitment tier” in the portal when you purchase.

How purchase works, as documented:

• Create a dedicated single-service resource on the standard pricing tier (disconnected containers need the separate Commitment tier disconnected containers tier and access).
• In the resource, open Commitment tier pricing, choose hosted web APIs and/or connected containers (on-premises containers that still connect for billing).
• Choose a tier and an auto-renewal option. Purchase enables immediately and charges a pro-rated amount for the rest of the current calendar month.
• Once you select Purchase you will be charged. Once purchased, the commitment plan is non-refundable.
• After the first pro-rated month, the charge hits on the first day of the calendar month.
• If usage exceeds the included quota, you pay overage at the rates shown for that tier.
• You cannot change the commitment plan for the current month. You can pick a different plan for the next calendar month; that bill occurs on the first day of that month.
• To stop, set auto-renewal to Do not auto-renew. The plan ends on the displayed end date, then the resource continues at Standard PAYG pricing. You have until midnight (UTC) on the last day of each month to end a plan and avoid the next month’s commitment charge.

Disconnected containers are a different commitment: calendar year, full price immediately, no mid-term plan change, extra units pro-rated for remaining days in the year, and extra quota purchased by raising unit count rather than a simple PAYG overage line. Do not mix disconnected-container rules with the monthly web/connected-container story on the exam.

Operating the two models as a leader

Roles: Cost Management Reader to see costs; Foundry User (recently renamed from Azure AI User; role IDs unchanged) to see Foundry usage context. Creating custom roles needs Owner at the scope. Foundry portal Operate and Monitor tiles show estimates that do not include discounts or contracted pricing and may omit some agent types; invoices and Cost Management remain the financial source of truth. Token charts can lag estimates because of ingestion delay.

Worked leader workflow from Microsoft’s cost article: estimate in the Azure pricing calculator → send representative test traffic → in Cost Management group by Resource then Meter → explain variance → set budgets after at least one billing-cycle trend. Recalibrate noisy alerts after you see a normal week, not on day one.

If the data-science team fine-tunes a model on PAYG, finance should see training, then a flat hosting burn, then inference spikes. If the contact-center team wants predictable Speech-to-text, finance should see a single-service Speech resource on a commitment, not a hope that the shared multi-service Foundry account “has a discount checkbox.”

AB-731 will not ask you for an unpublished dollar SKU. It will ask whether the organization is on PAYG versus commitment, whether the resource is single-service, and whether someone is watching overage, hosting, and Marketplace lines.