Evaluate models on the deployment type you will run in production, confirm quota before launch, plan spillover and a fallback route, and set identity, logs and budgets before the first production request. Each step cites Microsoft documentation and public accounts from teams running Foundry.
When you need this
- A feature or agent built on Foundry is moving from prototype to production traffic.
- The team is choosing between OpenAI, Claude and open models in Foundry for the same workload.
- A launch date is fixed, and nobody has checked the quota tier for the model and deployment type the launch needs.
Evaluation criteria
Model choice: evaluations on your task
Which Foundry model passes your own test set at the lowest cost per completed task?
Look forRun the same prompts on two or three models through the deployment type you will use. Compare model router against a fixed model on the same set before letting it pick per prompt.
Deployment type
Does the workload need Data Zone or regional processing?
Look forStart on Global Standard, which launches first at the lowest price. Move to Data Zone Standard for US, EU or Asia Pacific processing, or to Batch for work that can wait 24 hours at 50% off.
Capacity: quota, provisioned throughput and spillover
Is quota approved for each model and deployment type you will call?
Look forCheck the subscription's quota tier and file increases weeks ahead. Pair provisioned deployments with spillover to a standard deployment so peaks cost token rates and keep serving.
Hosted agents under load
Do your agents handle immediate follow-up requests and private networking?
Look forTest back-to-back turns: an East US 2 developer saw 409 conversation_locked when a follow-up arrived right after response.completed. Measure startup on private DNS paths before launch.
Fallback route
Where does traffic go during a regional incident?
Look forKeep a second region or route for each production model. Microsoft's review of Sep 29, 2026 records about six hours of intermittent errors in Sweden Central.
Questions to ask vendors
- Which quota tier is each production subscription in, and what did the last increase take?
- Which models on our roadmap lack a Data Zone deployment in our geography?
- Where does each production model fail over during a regional incident?
- Who owns the budget alert for each resource group?
Red flags
- A launch plan that assumes the quota a pilot used.
- Provisioned deployments with no spillover target.
- API keys in application settings when the app should use Entra ID tokens.
- One region and one deployment type with no fallback.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 3
- Microsoft Learn: Deployment types for Microsoft Foundry Models
- Microsoft Learn: Model router for Microsoft Foundry
- Microsoft Learn: Foundry Models quotas and limits
- Microsoft Learn: Provisioned throughput
- Microsoft Learn: Manage traffic with spillover for provisioned deployments
- Microsoft Learn: Authentication and authorization in Microsoft Foundry
- Azure status history: post-incident reviews
- Azure/azure-sdk-for-net issue 63892: hosted agent returns 409 conversation_locked after response.completed (Oct 15, 2026)
- Requesty: Operational metrics per provider, April 2026 (May 9, 2026)
- Azure/azure-sdk-for-python issue 49169 (Sep 24, 2026)