Microsoft Previews Azure API Management AI Gateway Tier

The preview supports models hosted through Microsoft Foundry, including OpenAI, Anthropic, Mistral, and other Foundry-hosted options. Microsoft also lists AWS Bedrock, Google Vertex AI, OpenAI, and Anthropic as supported model sources. InfoQ reports that OpenAI-compatible providers share one endpoint path, with request routing based on an exact match of the model field; published models therefore require unique names. Anthropic support uses a custom provider with Messages API passthrough, according to InfoQ.

Rather than configuring policies through the XML and expression-based approach familiar to API Management users, AI Gateway presents policy cards in the portal. Microsoft and InfoQ list controls covering token and request limits, quotas, content safety, and model fallback. The resource runs in a customer's Azure subscription, uses its Entra tenant, and exports telemetry to destinations controlled by that customer, Microsoft states.

For MCP and other tool backends, InfoQ reports that the gateway can federate a remote MCP server specified by URL, an OpenAPI specification, or a built-in connector. Each backend operation can be exposed as a tool. Available authentication options include no authentication, API keys, OAuth 2.0, and managed identities.

Microsoft describes a shared operating model in which a central platform group connects approved models and tools, publishes them, and maintains shared guardrails. Application teams can test published assets through a test console and create named runtime keys to call the gateway, according to the company.

The broader Azure API Management documentation frames AI gateway functions around securing, scaling, monitoring, and governing model, agent, and tool backends. It also identifies authentication, authorization, load balancing, logging, token-usage management, and quotas as common gateway functions across AI workloads.

For engineering organizations using several model providers, a gateway layer can reduce integration variance by centralizing provider-specific routing, identity controls, observability, and policy enforcement. Comparable multi-provider deployments still require careful evaluation of API-schema compatibility, model naming conventions, telemetry retention, and the behavior of fallback routes under quota or availability constraints.

Microsoft's public preview comes as MCP adoption expands the set of services that AI applications can invoke. In that environment, a centralized catalog and access-control layer can help platform teams apply conventional API-management disciplines to tool access, while application teams retain a self-service path to approved model and tool endpoints.