Model router frequently asked questions

Get direct answers to the questions developers, architects, and admins ask most about model router in Microsoft Foundry: how to start, which models it routes to, how quota and pricing work, and how to govern and monitor it.

  • Set model to auto to use Auto (model router without a deployment), which is in preview.
  • Deploy your own model router for production traffic, dedicated rate limits, and custom routing settings.
  • You pay a routing fee for each 1 million input tokens, plus the selected model's usage.
  • The model field in every response names the model that served the request.

If you can't find your answer here, see Troubleshoot model router or Foundry Tools support and help options. To get started, see What is model router? and Quickstart: Get started with model router.

Getting started

How do I start using model router without a deployment?

Use Auto (model router without a deployment), which is in preview. Set model to auto in the request body, and send the request to your Foundry resource or project endpoint, for example https://<resource>.services.ai.azure.com/openai/v1/responses. There's no /auto path and nothing to deploy. To send your first request, see Send your first request with Auto.

Do I need to enable anything to use Auto?

No. An endpoint and an API key are enough. With Microsoft Entra ID, the caller needs the Foundry User role (previously named Azure AI User) on the resource or project; see Role-based access control for Microsoft Foundry. Admins can turn off Auto or limit its models for a resource; see Configure Auto for your resource.

When should I use Auto instead of a model router deployment?

Use Auto for the fastest start, for prototypes, and for light traffic. Deploy your own model router for production traffic, dedicated and higher rate limits, a custom content filter, Data Zone Standard, or a routing mode and model subset for each deployment. For a side-by-side comparison, see Choose Auto or your own deployment.

Is model router available in Azure Government?

No. Model router isn't available in Azure Government. For the Foundry features that Azure Government offers, see Microsoft Foundry in Azure Government.

Can I keep data in a data zone, such as the EU?

Yes. Deploy model router with the Data Zone Standard deployment type, which processes data within the data zone. When your own deployments serve routed requests, it uses only your Data Zone Provisioned and Data Zone Standard deployments. Data Zone Standard is available in 23 of the 32 model router regions; see Supported regions. Auto always processes data as Global Standard, so use a deployment when data must stay in a data zone.

Models

Which models does model router route to?

As of November 11, 2026, model router version 2025-11-18 routes across 43 models from OpenAI, Anthropic, xAI, DeepSeek, Meta, and Fireworks. In each region, it routes only across the pool models available in that region, and each Claude model needs a deployment in your Foundry resource first. Fireworks models are used only when Fireworks is enabled for your Azure subscription. For the full list, with the exact name, version, and format values, see Supported models.

How quickly does model router add new models?

There's no published schedule for new models. New models join the pool of version 2025-11-18 over time, and retired models leave it. Deployments that use the full pool get new models automatically, but model subsets don't, so add new models to your subset yourself. What's new in model router lists each change.

Why doesn't model router route to Claude models?

Model router routes to a Claude model only when that model is deployed in the same Foundry resource as a Global Standard deployment of the version in the supported models table. The Foundry portal selects the Azure-hosted version 2 by default, and model router doesn't use it. If your deployment uses a model subset, the subset must include the Claude model. A Data Zone Standard model router deployment doesn't route to Claude models. Also check that the request's model names your model router deployment or auto, because a Claude model name in model doesn't go through model router. For the steps, see Add Claude models. For more fixes, see Troubleshoot Claude models.

Can I force one model for a request?

Yes, with per-request controls, which are in preview. Send the Foundry-Features: ModelRouterControls=V1Preview header, and set routing_config.models to a list with one model name, such as ["gpt-5.6-sol"]. Model router then pins the request to that model, as long as your deployment or Auto can already route to it. Per-request controls can narrow routing, but they can't add a model that your model subset or your admin's settings exclude. See Limit a request to specific models.

Can I add any model I've deployed to model router?

No. Model router routes only across the models in its pool, which you can narrow with a model subset but not extend. If you have your own deployment of a pool model, with the exact name and version, in the same Foundry resource, model router serves that model's requests from your deployment first: a provisioned throughput (PTU) deployment for OpenAI models, and then a pay-as-you-go deployment. See Supported models and Use your own deployments for routed requests.

Quota and capacity

How do I increase model router quota?

Submit the quota request form. Select Model Deployment as the quota type and Azure OpenAI as the model type, choose Global Standard or Data Zone Standard, and select ModelRouter as the model. Requests are reviewed manually. For more about model router quota, see Increase your quota.

Which rate limit applies to routed traffic?

Your model router deployment's tokens-per-minute (TPM) and requests-per-minute (RPM) limits apply to every request it routes, whichever model answers. You don't need quota for the underlying models, except for your Claude deployments and your own deployments that serve routed requests. When your own deployment serves a request, the request also uses that deployment's capacity: its provisioned throughput units (PTUs), or its quota and rate limits. See How model router quota works.

Does model router use my own model deployments?

Yes, in preview. After model router picks a model, it serves the request from your own deployment of that exact model name and version in the same Foundry resource, in this order: your provisioned throughput (PTU) deployment for OpenAI models, then your pay-as-you-go deployment (Standard, Global Standard, or Data Zone Standard) for every model provider in the pool, and then router-managed capacity. There's nothing to configure. See How model router chooses capacity.

Does model router use my provisioned throughput (PTU) deployments?

Yes, for OpenAI models, in preview. When model router selects an OpenAI model and the same Foundry resource has a Global Provisioned, Data Zone Provisioned, or Regional Provisioned deployment of that exact model name and version, that deployment serves the request first, on PTUs that you already pay for. If it returns a retryable error, such as HTTP 429 when it's fully used, model router moves to your pay-as-you-go deployment of the same model, if you have one, and then to router-managed capacity. There's nothing to configure. See Use your own deployments for routed requests.

What quota does Auto get?

In West US 3, Auto uses your subscription's Global Standard model router quota that isn't assigned to deployments. In every other model router region, Auto's limit is 1% of your subscription's model router quota: at Tier 1 (1,000,000 TPM), that's 10,000 TPM and 10 RPM, and it grows with your quota tier and with quota increases. The portal doesn't show this limit. For more capacity, request a quota increase or deploy model router; see Quota for Auto.

Pricing

How is model router billed?

You pay a model router routing fee for each 1 million input tokens, plus the selected model's usage at that model's rates. The routing fee applies to Auto, to model router deployments, and to requests that your own deployments serve. When your provisioned throughput (PTU) deployment serves a request, you pay only the routing fee, because you already pay for the PTUs. For current rates, see Azure OpenAI pricing. To track spend, see Track costs in Cost Management.

Is there a model router routing fee with AI Gateway?

No. There's no model router routing fee while automatic model selection in AI Gateway is in preview. See Use model router with AI Gateway.

Lifecycle

Is model router retiring in 2027?

No. Model router isn't retiring, but each model router version has its own retirement date: version 2025-11-18 retires May 20, 2027, and versions 2025-05-19 and 2025-08-07 retire January 27, 2027. Move deployments on the two older versions to 2025-11-18 before January 27, 2027. See Versions and lifecycle.

APIs and agents

Does model router work with the Responses API, including `previous_response_id`?

Yes. The Responses API works on the resource and project endpoints, including multi-turn conversations with previous_response_id or conversation, stored responses, background mode, streaming, and built-in tools. When a request continues a conversation, model router considers the earlier turns when it ranks models, and it doesn't switch models within a response. See Use the Responses API.

Does model router work with my agent framework?

Yes. Use a model router deployment as the model in Foundry Agent Service, Microsoft Agent Framework, LangGraph, the OpenAI Agents SDK, Google ADK, CrewAI, the Claude Agent SDK, or Claude Code. LangGraph and the OpenAI Agents SDK also work with Auto. Model router then picks a model for each step of your agent. For tested setup code, see Use model router with agents, agent frameworks, and coding tools.

Does model router work with GitHub Copilot, Codex, or OpenCode?

Yes. Add your model router deployment as a custom model: GitHub Copilot CLI and GitHub Copilot in Visual Studio Code use bring your own key (BYOK) with the Azure provider, Codex uses a model provider in ~/.codex/config.toml, and OpenCode uses a provider in opencode.json. Use a deployment rather than Auto, because coding agents send many large requests and need higher rate limits. See GitHub Copilot, Codex, and OpenCode.

Does model router work with Claude Code?

Yes, through the Messages endpoint. Set CLAUDE_CODE_USE_FOUNDRY=1 and ANTHROPIC_FOUNDRY_RESOURCE=<resource>, and set ANTHROPIC_DEFAULT_SONNET_MODEL, plus the Opus and Haiku variables, to your model router deployment name. Only the Claude models deployed in your resource answer, and a deployment works better than Auto because coding agents send many large requests and need higher rate limits. See Claude Agent SDK and Claude Code and Configure Claude Code for Microsoft Foundry.

Why is `temperature` ignored?

After model router selects a model, it drops optional parameters that the model doesn't support, such as temperature on reasoning models. To make every parameter mandatory instead, send the Foundry-Features: ModelRouterControls=V1Preview header and set routing_config.require_parameters to true (preview). Only models that support all your parameters are then eligible, and the request fails with HTTP 400 if none qualify. See How model router handles parameters.

Governance and data

My Azure Policy denies a model, but model router routes to it. Why?

Azure Policy is evaluated when a deployment is created or updated. It doesn't filter which models Auto or a deployment routes to at request time. To keep routing inside approved models, use a model subset on each deployment, and set properties.instant.routing.models for Auto. See How model router honors Azure Policy and Govern Auto.

Where is my data processed?

Model router doesn't store your prompts. With Global Standard, data can be processed in any geography where the selected model is deployed; with Data Zone Standard, it's processed within the data zone; and Auto always processes data as Global Standard. Claude models are provided by Anthropic; see Deploy and use Claude models in Microsoft Foundry. Model router uses Fireworks models only when an admin has enabled Fireworks for your Azure subscription and accepted its terms; see Enable Fireworks on Foundry. For details, see Data, privacy, and security for Models sold by Azure in Microsoft Foundry.

Observability

Which model served my request, and can I see why?

The model field in every response names the model that served the request, for example gpt-5.6-luna-2026-07-09. For more detail, send the Foundry-Features: ModelRouterControls=V1Preview header (preview); the response then includes model_selection_details, with the routing mode applied, the ordered model attempts that show any failover, and the session affinity decision. See Get per-request routing details. To learn how model router ranks models, see How a request is routed.

How do I show that model router saves money?

Before you adopt model router, evaluate it against a single-model deployment in the Foundry portal: each run reports quality scores, P50 and P95 latency, and, for Global Standard deployments, estimated cost. After you adopt it, the Monitor tab's Estimated cost: model router vs. single-model baselines chart shows what the same traffic would have cost on one model. Both are estimates, so confirm billed amounts in Cost Management. See Evaluate in the Foundry portal and Compare cost with single-model baselines.