Google Cloud API Gateway now routes AI traffic between models, and the whole feature is configured inside an OpenAPI document. In Model routing with Google Cloud API Gateway, Mak Ahmad and Sanjay Pujare describe a Public Preview in which the gateway “accepts OpenAI-compatible requests and dynamically routes them to Gemini, Claude, or OpenAI OSS-GPT.” Backends and routers are declared in a new x-google-api-management block in an OpenAPI 3.x spec, each path opts into a router with x-google-model-router, and the client sends a standard chat completions POST with a model name. The gateway transcodes the payload to the backend’s native schema, adds the platform token, and forwards it. The pitch is one sentence: “This gives you a single, stable endpoint for all your LLM traffic.”
There are no adoption or performance numbers, which is appropriate for a preview. The claims are architectural and they check out on their own terms. Because applications authenticate to the gateway rather than to the model providers, “client auth stays separate from backend LLM auth,” so backend credentials rotate without touching apps. The constraint worth reading twice is in a note: all backends behind one router must share the same host, so routing picks a different model and path on the agent platform host and does not route across hosts. That makes this a Vertex-hosted model switch with an OpenAI-shaped front door, not a general multi-vendor gateway, and the post is honest about the broader spectrum it sits inside, with Apigee for API management and Agent Gateway for agent governance.
The catalog carries the product as its own provider. The Google Cloud API Gateway provider page describes it as a fully managed, pay-per-use gateway for serverless workloads with authentication, API key validation, monitoring, logging, and tracing, which is the product the model router is bolted onto. It lists one API page, the Google Cloud API Gateway Projects API, the management surface where the API config in the post is created and deployed. The agentic access profile maps 11 operations, 6 of them acting.
The Kin Score is 49.5, developing band, the highest of the three providers profiled today. Discoverability carries it at 68.5, developer ergonomics at 58.3, contract quality at 55.8, access clarity at 55.3. Contract governance is 9.8. The Agent Readiness score is 19.8, agent-aware, and the unlit list is long: OpenAPI examples, error semantics, MCP server, agent skills, well-known catalog, idempotency. A feature that lives entirely inside OpenAPI extensions ships on a product whose own OpenAPI carries no examples and no described error semantics. The gateway is being sold as the door AI agents send their traffic through. By the catalog’s reading, an agent arriving at the gateway’s own API would find the door unlabelled.