Azure AI Foundry's Model Router: 28 Models, One Automated Decision

Instead of having a developer manually decide which model to use for each type of prompt, Azure's model router decides for you, in real time, based on how complex the request actually is.

What it is and how it works

Model router for Microsoft Foundry is a deployable chat model that automatically selects the best underlying model to answer a given prompt. It automatically routes inference requests to the best available model based on prompt complexity, latency targets, and cost efficiency -- without the developer having to make that decision manually on every call.

The decision flow, step by step

Your application sends an inference request to the Model Router deployment. The router analyzes the request (prompt complexity, token count, required capabilities). It selects the most appropriate model from the available pool. It's an ML model built specifically for this purpose, trained on hundreds of thousands of examples across diverse scenarios -- from simple prompts to complex agentic workflows.

Current coverage: 28 models and growing

The model router now supports 28 models. The 2025-11-18 release added support for seven new models: gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.3-chat, gpt-5.5, claude-opus-4-7, and grok-4.1-fast-reasoning -- notably, the router already crosses provider boundaries, not limiting itself to the GPT family.

Automatic cost optimization, no manual intervention

Simple prompts get routed to smaller, cheaper models; complex prompts go to more capable (and more expensive) ones -- automatically. The model router now also includes built-in automatic failover: if the chosen model fails or is unavailable, the system redirects the request without the application having to handle that error case explicitly.

When to use the router vs. choosing the model manually

The router makes sense when your application receives a high volume of prompts with highly variable complexity and you don't want to maintain your own routing logic. For applications with a predictable, homogeneous usage pattern (for example, a single well-defined use case like ticket classification), manually choosing and pinning the model usually gives more predictable cost and behavior than delegating the decision to an automatic system, even at the cost of losing dynamic optimization.