Veo 3.1 GA (`veo-3.1-fast-generate-001`) fails via google-genai (routes to v1beta1)

[thedavgar]

Environment details

• Programming language: Python
• OS: Windows 11 (also reproducible inside devcontainer)
• Container: python:3.11.8-slim
• Language runtime version: Python 3.11.8
• Package version: google-genai==1.64.0

pip show google-genai output:

Name: google-genai
Version: 1.64.0
Location: /usr/local/lib/python3.11/site-packages

---

Description

We are attempting to use Veo 3.1 via Vertex AI (vertexai=True) in the google-genai SDK.

Calling Veo 3.1 Fast GA model using client.aio.models.generate_videos() with Vertex AI enabled returns:

400 INVALID_ARGUMENT
reason: RESOURCE_PROJECT_INVALID
method: google.cloud.aiplatform.v1beta1.PredictionService.PredictLongRunning

However, calling the same model via direct REST to the documented Vertex AI v1 endpoint succeeds.

This suggests that when vertexai=True is set, the SDK is routing Veo GA calls to the v1beta1 PredictionService endpoint instead of the required v1 endpoint.

---

Steps to reproduce

1. Initialize client

from google import genai
from google.genai import types

client = genai.Client(
    vertexai=True,
    project="PROJECT_ID",
    location="us-central1",
)

2. Call Veo 3.1 Fast GA

operation = await client.aio.models.generate_videos(
    model="veo-3.1-fast-generate-001",
    prompt="A cinematic drone shot over ocean cliffs at golden hour",
    config=types.GenerateVideosConfig(
        number_of_videos=1,
        duration_seconds=4,
        aspect_ratio="16:9",
    ),
)

3. Observe error

Full error:

ClientError: 400 INVALID_ARGUMENT. {
  'error': {
    'code': 400,
    'message': 'Invalid resource field value in the request.',
    'status': 'INVALID_ARGUMENT',
    'details': [{
      '@type': 'type.googleapis.com/google.rpc.ErrorInfo',
      'reason': 'RESOURCE_PROJECT_INVALID',
      'domain': 'googleapis.com',
      'metadata': {
        'method': 'google.cloud.aiplatform.v1beta1.PredictionService.PredictLongRunning',
        'service': 'aiplatform.googleapis.com'
      }
    }]
  }
}

---

Control test (works)

Calling the documented REST endpoint directly works:

POST https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/veo-3.1-fast-generate-001:predictLongRunning

Returns:

200 OK
{
  "name": "projects/.../operations/..."
}

This confirms:

• Project is enabled for Veo
• Billing is enabled
• Region is correct
• IAM permissions are correct
• Model access is valid

---

Expected behavior

The SDK should route Veo 3.1 GA model requests to:

v1 PredictionService.PredictLongRunning

matching the documented REST behavior.

---

Actual behavior

SDK routes to:

google.cloud.aiplatform.v1beta1.PredictionService.PredictLongRunning

which results in RESOURCE_PROJECT_INVALID.

---

Additional Notes

• Gemini models work correctly via the same client configuration.
• Issue reproduces consistently in both Windows 11 and Docker (python:3.11.8-slim).
• Raw REST calls confirm product entitlement is correct.
• Appears specific to Veo GA long-running video generation when using google-genai with vertexai=True.

---

Hypothesis

It appears that the SDK may still be routing Veo long-running publisher model calls through the v1beta1 PredictionService when vertexai=True is enabled, while the GA Veo model requires the v1 endpoint.

If so, the endpoint mapping for long-running video generation models may need to be updated internally within the SDK.