Key findings
- Foundry lists more than 10,000 models. Azure OpenAI models are sold directly by Azure, while Claude comes in two versions, hosted on Azure or on Anthropic infrastructure.
- Global Standard launches first at the lowest price; Data Zone keeps processing in the US, EU or Asia Pacific, and Batch takes 50% off with a 24-hour window.
- Quota follows seven tiers set per subscription from usage, payment history and agreement status.
- Requesty's gateway saw 78% of Azure requests succeed in April 2026, and Microsoft's review of Sep 29, 2026 records about six hours of errors in Sweden Central.
Microsoft Foundry, formerly Azure AI Foundry, groups agents, models and tools under one Azure resource with Entra ID, role-based access control and Azure Policy. Underneath, each model family keeps its own deployment types, quota and data terms, so buying Foundry means checking every model separately. Public accounts and Microsoft's own incident reviews show where that matters.
What Foundry bundles
Microsoft describes Foundry as one management grouping for agents, models and tools, with tracing, monitoring, evaluations and one set of access controls, networking and policies. Foundry Agent Service runs declarative prompt agents and hosted agents that run the team's own code, and agents reach business systems through MCP, OpenAPI tools and Logic Apps connectors.
Azure OpenAI is now sold inside Foundry. An Azure OpenAI resource can be upgraded to a Foundry resource as an opt-in step that can be rolled back, though resources with private endpoints must delete and recreate them or use Bicep or Terraform.
How each model family is sold
Prices come from Azure's retail meters per model and deployment type. GPT-6.1 Sol on Global Standard in East US 2 lists US$2 input and US$10 output per million tokens, effective Oct 1, 2026, and Llama 4 Maverick on Global Standard lists US$0.25 and US$1.00. Claude requires a paid Azure subscription with an active pay-as-you-go billing account.
Rajesh Beri's October 2026 analysis rates Foundry strongest for GPT models and weakest for Claude among the three clouds, and argues that committed-spend drawdown should decide the platform more than per-token price.
Quotas, agents and reliability in public accounts
Quota is the most common complaint. Hosted agents show early defects: a developer in East US 2 found that a follow-up request sent right after response.completed failed with 409 conversation_locked, because the conversation store write had not finished.
Teams still pick Foundry for compliance and procurement. One engineer describes a 100-plus developer organization that runs Claude Code against models in its own Foundry, signing in with az login and no keys. Reliability numbers are thinner: Requesty measured 78% success on Azure in April 2026, and one Hacker News commenter cites 87% uptime for Azure US without naming the source.
Methodology
GILD built this piece from Microsoft Learn pages on Foundry, deployment types, partner models, quotas, provisioned throughput, spillover, data privacy and the Azure OpenAI upgrade, the Azure retail prices API and Azure status history, all checked Oct 28, 2026. Prices are retail list prices in US dollars per million tokens for East US 2. Practitioner accounts are public posts dated 2026, each linked in Sources.
Products in this piece
Sources
- Official documentation
- 9
- Public source
- 5
- Microsoft Learn: What is Microsoft Foundry?
- Microsoft Learn: Foundry Models from partners and community
- Microsoft Learn: Foundry Models sold directly by Azure
- Microsoft Learn: Deployment types for Microsoft Foundry Models
- Microsoft Learn: Foundry Models quotas and limits
- Microsoft Learn: Upgrade from Azure OpenAI to Microsoft Foundry
- Azure retail prices: Foundry Models GPT-6 meters, East US 2
- Azure retail prices: Llama 4 Maverick meters, East US 2
- Azure status history: post-incident reviews
- Requesty: Operational metrics per provider, April 2026 (May 9, 2026)
- Rajesh Beri: Bedrock vs Vertex AI vs Microsoft Foundry, model platform commit drawdown (Oct 25, 2026)
- Azure/azure-sdk-for-net issue 63892: hosted agent returns 409 conversation_locked after response.completed (Oct 15, 2026)
- Hacker News comment on running Claude Code against Foundry with az login (Oct 15, 2026)
- Hacker News comment on Azure-hosted GPT-5.6 Sol uptime and latency (Aug 18, 2026)