Foundry Model Router Now Covers 28 Regions

Microsoft recently expanded its model router in Foundry Models from two regions to 28 for global standard deployments and 21 for data zone deployments, and refreshed the pool of models it selects from. The expansion aims to improve compliance with regulatory requirements by keeping inference requests within specific geographic boundaries.

The model pool now includes Anthropic Claude Opus 4.8 and the GPT-5.6 family, while removing older models like gpt-5-chat and DeepSeek-V3.1 that have reached their end of life. This change is automatic for teams using the default configuration, but teams that have configured a custom model subset will need to manually update their settings.

Sanjeev Jagtap, who announced the update, emphasized that the changes occur automatically, and teams do not need to redeploy the model router to receive the update. Christos Panagiotidis, an Azure MVP and cloud engineer, noted that while the endpoint remains stable, the introduction of new models can change the application’s business outcome.

Panagiotidis explained that a new model in the pool can affect answer style, tool-selection behavior, and latency distribution, even if the response schema remains identical. This means that teams need to carefully evaluate the impact of the updated model pool on their specific use cases.

The model router has three routing modes: Balanced, Quality, and Cost. The Balanced mode is the default and optimizes cost while maintaining quality. Teams can restrict routing to a chosen subset of models, but this requires explicit configuration. The update is opt-out, and teams that have constrained their pool are insulated from the refresh.

Changes to the routing mode or subset take up to five minutes to apply. Additionally, the effective context window equals the smallest underlying model’s window, which means that adding a smaller model to the pool can lower the ceiling for every request routed through it.

The model router has some limitations, including text-only routing decisions, which means that vision inputs are accepted but images do not influence model selection. Audio inputs are also unsupported. Furthermore, the router bills its own input prompt on top of the underlying model’s cost, which can affect cost savings claims.

Anthropic models, such as Claude Opus 4.8, require separate deployment to the same Foundry account with a matching SKU before the router can select them. This prerequisite is noted in the announcement and documentation, but it may not be immediately apparent to teams using the model router.

Teams will need to consider the implications of the changes on their specific use cases as they work with the updated model pool. They may involve treating pool refreshes as managed dependency updates, capturing the selected model for every evaluated request, and comparing against previous periods to ensure acceptable outcomes.

It is essential to evaluate the impact of the updated model pool.

The expansion of the model router to 28 regions and the refresh of the model pool aim to improve compliance and flexibility for teams using Foundry Models. While the update brings new opportunities, it also requires careful consideration of the limitations and potential impacts on application performance, such as those related to model updates.