Request access: gemini-3.8-live on Vertex AI BidiGenerateContent (1008/no access)

Google Cloud Build with AI Generative AI & Foundational Models
1

Hello Google Cloud team,

We would like to request Vertex AI Live API access to the exact model gemini-3.8-live for evaluation of our Swedish AI receptionist, Hejvinter. If this requires preview allowlisting, please route this request to the relevant team or provide the official application process.

Our standalone demo runs on Cloud Run and uses Vertex AI with Application Default Credentials, not a Google AI Studio API key. Billing is enabled, the Vertex AI API is enabled, and the runtime service account has the Vertex AI User role.

On September 18, 2026, attempts to open a BidiGenerateContent session failed with WebSocket close code 1008 and a publisher-model-not-found/no-access message. We tested gemini-3.8-live and gemini-3.8-live@default, including regional attempts in europe-west4 and us-central1; no Live session reached setup completion. The model catalogue listing alone does not establish Live API availability.

Could you please:

  1. Confirm whether this exact model currently supports Vertex AI bidirectional streaming, and the required model ID, API version and supported region.
  2. Enable our project if an allowlist is required, or direct us to the official access request form/team. We can provide our project ID and number privately to Google staff.
  3. Confirm whether usage through Vertex AI is eligible for our Google for Startups Cloud credits, subject to our credit terms.

Our preferred inference region is europe-west4. We can use us-central1 for this evaluation if that is the supported option.

Thank you,
Nima
Autorank / Hejvinter

2

Hi Nima,

Regarding the WebSocket 1008 / model-not-found error on Vertex AI:

1. Model Availability (Root Cause #1):
gemini-3.8-live is currently restricted to private preview / allowlist and is not accessible on public Vertex AI endpoints. As confirmed directly in Google Cloud Agent Platform / Vertex AI, the active and available 3.x models are:

  • gemini-3.5-flash (General Availability / Production)
  • gemini-3.1-pro (preview)

If your project is not allowlisted for the private 3.8-live preview, the WebSocket handshake will immediately be rejected with code 1008 (no access / model not found).

2. Endpoint Location (Root Cause #2):
In your post, you mentioned testing regional endpoints like europe-west4 and us-central1. According to the official Google Cloud ADK (Agent Development Kit) architecture specifications:

Gemini 3-series models are served exclusively through the global location endpoint.

Calling regional endpoints (like us-central1 or europe-west4) for Gemini 3.x models triggers a 404 / 1008 error.

3. Dual Operability (Local PC & GCP Cloud Run):
It is 100% viable and verified to run this agent setup seamlessly in both environments:

  • Locally on your PC: Using Python with local ADC (gcloud auth application-default login), allowing you to test and validate the full conversational loop on localhost before deploying.
  • In GCP (Cloud Run / Agent Platform): Running with default service account credentials.

4. Working Client Setup:
Ensure your client explicitly pins the location to global:

from google.genai import Client

# Explicitly use location="global"
client = Client(vertexai=True, project="YOUR_PROJECT_ID", location="global")
3

Forums might not help you here. Please create a ticket here : https://console.cloud.google.com/support