Google API Gateway adds streaming in public preview
API Gateway now carries SSE, WebSockets, gRPC and chunked responses in Public Preview without changes to existing API specifications.
On Oct 22, 2026, Google put streaming support in API Gateway into Public Preview. Per the announcement on the Google Developers Blog, a gateway can now carry long-lived and bidirectional streams while keeping the authentication, rate limiting and observability it already applies to ordinary requests. Google says the API specification you already maintain describes the streaming version of your API too.
That removes a split that teams had been living with. Until now the choice was to put API Gateway in front of APIs and lose streaming, or to route streaming endpoints around it and lose key enforcement, JWT validation and logging. The announcement describes the result as a split estate, with governed traffic on one path and ungoverned streams on another.
What the preview actually streams
Google lists four patterns. Server-Sent Events, which the post calls the standard transport for pushing model output to a browser or client SDK. Chunked transfer encoding and HTTP response streaming, including newline-delimited JSON, where a backend's output is passed through as it is produced rather than buffered until the response finishes. WebSockets, for full-duplex chat, collaboration, dashboards and notifications. And gRPC server-side plus bidirectional streaming, which the announcement says could not previously be proxied through the gateway.
Streams can stay open for up to one hour. Google notes that clients should implement reconnection logic so a session survives the end of a stream.
How you switch it on
Streaming is enabled when the gateway is created, with an --enable-streaming flag added to the gcloud api-gateway gateways create command. Existing gateways cannot be converted in place, so testing the preview means creating a new one. Google states that at general availability, streaming is planned to become the default behaviour with no opt-in, and still with no changes to API config.
Everything already in use is said to keep working: API key enforcement, JWT validation including Firebase, Auth0, Okta and Google ID tokens, service account authentication to backends, Cloud Logging and Cloud Monitoring, path templating and request routing.
A hostname you can write down first
Gateways also get a deterministic default hostname of the form GATEWAY_ID-PROJECT_NUMBER.REGION.gateway.dev, with streaming-test-12345.us-central1.gateway.dev as the example given. The previous hostname included a generated hash, so nobody could know the URL until the resource existed. That blocked putting the endpoint into client config, Terraform modules, CI pipelines, firewall allowlists or partner docs, and tearing an environment down and rebuilding it changed the URL. The new form is derived from values already chosen, and rebuilding the same gateway in the same project and region produces the same address.
Why AI traffic pushed this
The post argues that streaming is the expected interaction pattern for model-backed features, not a niche case: users expect tokens to appear as they are generated, agents report intermediate steps, and tool-calling protocols and remote MCP servers rely on connections that stay open. Model routing in API Gateway lets you expose one OpenAI-compatible endpoint and dispatch requests to Gemini, Claude or OpenAI gpt-oss models in model garden in Gemini platform based on the model field in the payload, with responses streamed back. Google places API Gateway as the lightweight entry point alongside Apigee and Agent Gateway, and says the gateways can be combined, including running API Gateway behind Agent Gateway.
Limits and open questions
Some things are missing at the initial Preview launch and described as coming during the Preview period: custom domains via Serverless NEGs and an external Application Load Balancer, streaming enablement through the Cloud console, and Terraform support. Latency metrics for streaming connections measure the life of the stream rather than time to first byte, so long-lived connections will change the shape of latency graphs.
The source gives no GA date, no pricing for streaming traffic, and does not say how per-stream duration interacts with rate limits or quotas. It also does not state whether model routing and MCP support are themselves generally available.
What this suggests for app builders
This is inference rather than anything the source confirms: a gateway-level streaming story mainly matters once an app's AI features leave the prototype stage and need per-consumer keys, quotas and logs around a model endpoint. Apps talking to a single model directly on the client will not notice much. If you are building on Draftship, the guide to adding AI to your app covers the in-app side of that work, while Draftship itself is an AI app builder with built-in hosting, a database and sign-in, and has no connection to this announcement.
Primary sources
- Streaming comes to API Gateway: real-time and AI-ready APIs, without changing your spec · Google Developers Blog
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.