What changed in Azure AI Foundry Model Router and why does it matter for production workloads?

New regions and a changed model list

The router now runs in 28 global standard regions and 21 data zone regions. Organizations that must keep inference inside specific borders can use intelligent routing without moving data.

The model pool grew and shrank at the same time. Anthropic Claude Opus 4.8 joined the lineup for complex reasoning and long-form generation. The GPT-5.6 family arrived alongside it. Four models left: gpt-5-chat, gpt-5.2-chat, gpt-5.3-chat, and Deepseek-V3.1 reached end of life and were deprecated in Foundry. The endpoint stays stable. Teams do not redeploy to receive the new pool. Applications keep the same integration while the router evaluates requests against the current supported list. Anthropic models still need a separate deployment before the router can send traffic to them. The blog notes that teams should continue monitoring routing traces and application outcomes to confirm quality, cost, latency, and governance requirements are met.

How the router turns model choice into a loop

Microsoft calls this pattern hill climbing. Each routing decision is a micro-optimization. Every response includes a model field showing which underlying model was selected. The climb leaves a complete auditable trail.

What teams still have to do themselves

Evaluation remains a human job. Teams must supply representative traffic for A/B tests. They must examine telemetry for decomposition insights. They must define the objectives that guide each micro-optimization.

Three patterns support the loop. A/B testing compares two router configurations against the same workload. Teams measure quality, cost, latency, and governance outcomes before promoting a winner to production. Model decomposition uses the router as a diagnostic tool. Routing telemetry reveals how requests naturally separate into task classes. Simple retrieval and classification may route to smaller models. Reasoning and agentic workflows may require more capable models. The goal is not to pick a winner but to understand workload structure and uncover optimization opportunities. Route continuously keeps the router in production for per-request selection. As the pool, regions, and platform capabilities evolve, the same endpoint keeps working. Model selection becomes an ongoing process rather than a project repeated every time the landscape changes.

The blog points to several resources. Model Mondays with Sanjeev Jagtap and Lee Stott covers evaluations and hill climbing. The Model Router Agents Lab teaches routing, retrieval, web search, tool calling, and multi-agent patterns. The Auto Evaluation Toolkit compares the router against baseline models using custom prompts. The Model Mastery workshop tests one optimization lever at a time. The Model Releases repository tracks new capabilities and offers focused notebooks. None of these run automatically. Teams decide which to use and when. The router provides the stable endpoint and the telemetry. The optimization loop remains a human responsibility.