Quick answer
Google’s launch example proves that one API Gateway model router can map virtual names to Gemini, Anthropic Claude, and OpenAI OSS-GPT targets hosted through Vertex AI MaaS—provided the referenced backends share one aiplatform.googleapis.com hostname.
That does not make their native APIs identical. The gateway transcodes an OpenAI-compatible request to the configured target schema, and each backend keeps a model-specific Vertex AI path.
Published example pattern
| Family | Example backend shape in Google’s launch | Route check |
|---|---|---|
| Gemini | publisher Google model :generateContent | Exact Gemini model ID, location, modality, quota |
| Claude | publisher Anthropic model :rawPredict | Exact Claude MaaS ID, access, region, request behavior |
| OpenAI OSS-GPT | Vertex AI MaaS OpenAI chat-completions endpoint | Exact OSS model target, compatibility, quota |
The launch happens to use then-current Gemini 3.5 Flash-Lite, Claude Opus 4.7, and GPT-OSS-120B identifiers as examples. Treat model IDs as live deployment inputs, not permanent route names.
Design virtual names deliberately
Avoid exposing a vague name such as best unless the application owns and logs an evaluated selection policy. Prefer stable workload names such as support-fast or analysis-reviewed, then map them to exact backends in a versioned config. When a backend changes, rerun the workload evaluation and preserve the old config for rollback.
Test each target for:
- message, role, system, and content behavior;
- tools, structured output, safety settings, and unsupported parameters;
- streaming and error response parity;
- requested versus resolved model logging;
- token usage, quotas, latency, price, and location;
- refusal and final-answer differences.
An OpenAI-compatible request shape does not guarantee behavioral or feature parity. Use the OpenAPI guide to keep target paths explicit and the rate and token guide for operational controls.
Frequently asked questions
Can API Gateway route to Gemini, Claude, and OpenAI OSS-GPT?
Yes. Google’s Public Preview launch example configures Gemini, Anthropic Claude, and OpenAI OSS-GPT MaaS targets hosted through Vertex AI.
Do all three model families use the same backend path?
No. Google’s example uses publisher- and model-specific paths, including generateContent for Gemini, rawPredict for Claude, and an OpenAI-compatible chat-completions path for OSS-GPT.
Can the router call models outside Vertex AI?
The Preview documentation limits model routing to configured pre-deployed MaaS models hosted on Vertex AI Model Garden, and one router’s backends must share one Vertex hostname.
Official sources
- Google Developers Blog launch and model example
- API Gateway model routing overview
- Use open models with Vertex AI MaaS
Source check: August 5, 2026. Verify current model IDs, endpoints, access, regions, quotas, prices, features, and terms.