Microsoft expanded Foundry's model router from two regions to 28 for global standard and 21 for data zone deployments, while adding Claude Opus 4.8 and GPT-5.6 and removing four deprecated models. Default deployments receive pool changes automatically; configured subsets exclude new models until added. The effective context window equals the smallest model in the pool.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Microsoft has expanded the Foundry model router from 2 regions to 28, with automatic model pool refreshes, but the default auto-update may introduce behavioral changes, and the subset configuration in the documentation is the real escape hatch.
Microsoft recently expanded the Foundry model router from two regions (East US 2 and Sweden Central) to 28 global standard deployments and 21 data-zone deployments, and refreshed the model pool. Added Anthropic Claude Opus 4.8 and GPT-5.6 series, while removing gpt-5-chat, gpt-5.2-chat, gpt-5.3-chat, and DeepSeek-V3.1, as these models have reached end-of-life.
Teams with default configuration automatically receive updates without redeployment, while teams with configured model subsets do not. This difference is described differently in the announcement and documentation: the announcement emphasizes automatic delivery, while the documentation provides subset configuration as an escape hatch.
Azure MVP Christos Panagiotidis pointed out that API stability and behavioral stability are two different things. New models may change response style, tool selection behavior, structured output reliability, latency distribution, token usage, refusal behavior, and failure modes. Response patterns may remain the same, but business outcomes may change.
The default deployment uses Balanced mode, covering the full supported set, so workloads now have two candidate models that have never been evaluated, four models that may have been used are gone, and there is no deployment or version upgrade. The subset configuration in the documentation can exclude new models, but most teams will not modify this setting.
The router has three modes: Balanced (default, optimizes cost while maintaining quality), Quality (for critical work, such as legal review, medical summarization, and complex reasoning), and Cost (for high-volume classification and simple Q&A). Mode or subset changes take up to five minutes.
Microsoft previously published constraints stating that the effective context window equals the smallest model's window in the pool, and oversized prompts only succeed if the router happens to select a model that can handle them. Therefore, adding smaller models to the pool lowers the limit for all requests. Additionally, routing decisions are based solely on text; visual inputs are accepted but do not influence model selection, and audio is not supported. The router also charges its own input prompt fee on top of the underlying model costs.
Anthropic models have prerequisites: Claude models must be separately deployed to the same Foundry account and match SKUs for the router to select them; referencing them in a subset without deployment results in an InvalidResourceProperties error. Claude Opus 4.8 being added to the supported list does not mean direct access.
Microsoft frames the regional expansion as a compliance requirement, requiring inference requests to stay within specific geographic boundaries. However, the materials do not explain the interaction between data-zone constraints and pool selection when candidate models are unavailable. The router follows built-in Azure Policy, enforced at deployment time, and the allowed publisher list must include Microsoft and the publisher of each model in the pool.
This article is based on InfoQ's report, with information from Microsoft announcements and documentation, as well as comments from an Azure MVP. Some details (such as model pool changes, region counts) are as stated by the source and have not been independently verified.
While the Foundry model router's automatic updates simplify operations, they may introduce behavioral changes. Teams should evaluate subset configuration and context window limits to maintain business stability.
Primary report
Primary source