AI Infrastructure
Microsoft's Model Router Now Spans 28 Regions With Updated AI Model Support
Microsoft expanded its Foundry model router across global infrastructure while adding new AI models, but automatic updates pose challenges for production workloads.
InfoQ
Microsoft has dramatically scaled its Foundry model router infrastructure, extending it from a two-region footprint to 28 regions for standard deployments and 21 for data-residency-constrained environments. The expansion also introduced fresh model options while retiring others, creating both opportunities and operational complexities.
What Changed#
The router now includes Anthropic's Claude Opus 4.8 and GPT-5.6 family models, while retiring gpt-5-chat, gpt-5.2-chat, gpt-5.3-chat, and DeepSeek-V3.1 after they reached end of life. Teams using the default configuration receive these updates automatically without redeployment, whereas those with custom model selections remain unaffected until they opt in.
The Automatic Update Challenge#
The announcement emphasizes that API stability persists—the endpoint remains unchanged—but this masks a deeper shift. A new model in the pool can alter response style, tool selection, structured output reliability, latency, token consumption, refusal behavior, and failure patterns. As cloud engineer Christos Panagiotidis notes, "API stability and behavioral stability are different things." The response schema may stay identical while application outcomes diverge.
Key Constraints and Trade-offs#
The router offers three modes—Balanced (default, optimizing cost with quality), Quality (for critical work like legal review and medical summaries), and Cost (for high-volume classification)—none mentioned in the announcement. Additional operational limits warrant careful attention:
- Context windows: The effective context size equals the smallest model in the active pool. Adding smaller models lowers the ceiling for all requests.
- Vision and audio: Routing decisions are text-only; images don't influence model selection, and audio remains unsupported.
- Billing: The router bills its own input prompt separately, meaning any cost savings include this markup.
- Claude prerequisite: Anthropic models must be pre-deployed separately with matching SKU configurations before the router can reach them, even when listed as supported.
