Skip to main content

Decision brief

Microsoft Foundry Review: Buying Models and Agents Through Azure

Foundry, formerly Azure AI Foundry, sells OpenAI, Anthropic, Meta and open models and runs agents inside an Azure subscription. Check deployment types, quota tiers and each model's data terms before production.

Reading time
4 min read
Sources
17 cited
In this piece
Microsoft Foundry−26 Negative

GILD verdict

Standardize on Microsoft Foundry when your organization runs on Azure, signs in with Microsoft Entra ID, and wants OpenAI's models and Claude under one Azure agreement.

Start each model on Global Standard unless residency rules require Data Zone or regional processing. Before production traffic, confirm the quota tier for every model, and keep a fallback route for the regional incidents Microsoft's own reviews record.

Microsoft Foundry

AI Model API by Microsoft

−26 Negative
Product page

Fit and limits

Who it fits

  • Your organization runs on Azure, signs in with Entra ID, and counts AI spend toward an Azure commitment.
  • You need OpenAI's GPT models sold directly by Azure, next to Claude, Llama, DeepSeek and Microsoft's own models.
  • Processing must stay in the US, the EU or Asia Pacific: Data Zone deployments keep it inside that zone.
  • You want agents and their governance in the same Azure resource: Foundry Agent Service, MCP and OpenAPI tools, Logic Apps connectors, built-in evaluators and Purview audit.

Where it falls short

  • Claude in Foundry requires a paid Azure subscription with an active pay-as-you-go billing account; Cloud Solution Provider subscriptions and Enterprise Accounts in South Korea are not supported.
  • Quota tiers start from usage and agreement status, so a new subscription can begin with little or no quota for the model it needs.
  • Requesty's gateway saw 78% of Azure requests succeed in April 2026, the lowest of the major routes it measured.
  • Microsoft's post-incident review records about six hours of intermittent failures and HTTP 500 and 503 errors in Sweden Central on Sep 29, 2026, across Azure OpenAI, Foundry Agent Service and Foundry Models.

Implementation reality

  1. Choose models and deployment types

    Run your own test set on two or three Foundry models. Start on Global Standard, which launches first with the lowest price and broadest region coverage, and move to Data Zone or regional deployments only for residency, reserved throughput or batch. Model router can choose the model per prompt when the routing settings fit your quality and cost targets.

  2. Plan quota and capacity

    Quota sits in seven tiers, Free and Tiers 1 to 6, evaluated per subscription. Request more through the quota form before launch. For steady traffic, buy provisioned throughput units per hour or by reservation, and turn on spillover so excess traffic moves to a matching standard deployment at token rates.

  3. Set identity and data terms

    Use Microsoft Entra ID with keyless, token-based access for production and keep API keys out of it. Prompts and completions for models sold by Azure stay out of OpenAI's and other model vendors' reach and are not used to train foundation models without permission. Claude has its own terms, with Anthropic as the data processor.

  4. Log and cap spend

    Send resource logs to Azure Monitor and turn on Purview Audit, which records prompts and responses from Foundry interactions in the unified audit log. Set Cost Management budgets per subscription and resource group before launch.

What to verify before you buy

  1. Foundry quota comes in seven tiers, Free and Tiers 1 to 6, set per subscription from usage, payment history and the Azure offer.
  2. Prompts and completions for models sold by Azure are not available to OpenAI or other model vendors and are not used to train foundation models without permission.
  3. Before Microsoft Purview can manage Foundry AI interactions, you enable Microsoft Purview Data Security in the Microsoft Foundry Control Plane or in Microsoft Defender for Cloud.
  4. Prompts and responses are then captured in the unified audit log, and Microsoft Purview Audit includes Microsoft Foundry data as part of the Microsoft Purview license.
  5. Your organization's own Purview retention policies can then automatically retain or delete user prompts and responses for AI apps, with the Enterprise AI apps option selected and a collection policy for prompts and responses, and managing these policies requires pay-as-you-go billing in your organization.
  6. Microsoft lists Microsoft Foundry among its AI services in scope for ISO/IEC 42001 certification, the standard for an Artificial Intelligence Management System.
  7. For Models sold by Azure, which include Azure OpenAI models, a sample of prompts and completions selected for review may be stored for human review in an abuse monitoring data store in the Azure geography where the customer's Foundry resource is deployed. If the customer has been approved for modified abuse monitoring, this data storage and human review is not performed. Models sold by Azure in preview may employ different privacy practices.
  8. Model router is a trained language model that routes your prompts in real time to the most suitable large language model, and you deploy it like any other Foundry model. Its current version, 2025-11-18, supports Global Standard and Data Zone Standard deployments.

What to measure in your pilot

  • GILD found no published time for a quota tier increase; file the request at least four weeks before launch and record how long Microsoft takes.
  • Requesty's April 2026 figure covers Azure traffic as a whole. Measure your own success rate per deployment type for two weeks before committing to provisioned throughput.
  • Recheck Claude's deployment types in Foundry monthly; Microsoft documents them per model, and the EU data zone is absent for some releases.

Sources

Official documentation
17
  1. Microsoft Learn: What is Microsoft Foundry?Official documentationlearn.microsoft.com
  2. Azure: Microsoft FoundryOfficial documentationazure.microsoft.com
  3. Azure: Microsoft Foundry pricingOfficial documentationazure.microsoft.com
  4. Azure retail prices: Foundry Models GPT-6 meters, East US 2Official documentationprices.azure.com
  5. Microsoft Learn: Deployment types for Microsoft Foundry ModelsOfficial documentationlearn.microsoft.com
  6. Microsoft Learn: Foundry Models from partners and communityOfficial documentationlearn.microsoft.com
  7. Microsoft Learn: Foundry Models sold directly by AzureOfficial documentationlearn.microsoft.com
  8. Microsoft Learn: Model router for Microsoft FoundryOfficial documentationlearn.microsoft.com
  9. Microsoft Learn: Foundry Models quotas and limitsOfficial documentationlearn.microsoft.com
  10. Microsoft Learn: Provisioned throughputOfficial documentationlearn.microsoft.com
  11. Microsoft Learn: Manage traffic with spillover for provisioned deploymentsOfficial documentationlearn.microsoft.com
  12. Microsoft Learn: Data, privacy, and security for Foundry ModelsOfficial documentationlearn.microsoft.com
  13. Microsoft Learn: Authentication and authorization in Microsoft FoundryOfficial documentationlearn.microsoft.com
  14. Microsoft Learn: Microsoft Purview for Microsoft FoundryOfficial documentationlearn.microsoft.com
  15. Azure status history: post-incident reviewsOfficial documentationazure.status.microsoft
  16. Microsoft Learn: Use Microsoft Purview to manage data security & compliance for Microsoft FoundryOfficial documentationlearn.microsoft.com
  17. Microsoft Learn: ISO/IEC 42001:2023 compliance offeringOfficial documentationlearn.microsoft.com

Keep reading

  1. Buyer guideMicrosoft Foundry Production PlaybookA checklist for taking a model or agent on Foundry from prototype to live traffic: deployment type, quota tier, hosted agent behavior and spend control.
  2. Buyer guideFoundry Data, Identity & Residency ChecklistWhich deployment type keeps processing in a geography, how each model family's data terms differ, and how Entra ID, audit logs and compliance scope apply.
  3. ResearchMicrosoft Foundry in 2026: One Azure Resource for Models and AgentsWhat Foundry bundles after the Azure AI Foundry rename, how each model family is sold and billed, and what public accounts report about quotas, hosted agents and reliability.
Read the research