Intermittent 403 and Quota Errors with Gemini 3 Series on Vertex AI (Paid Tier, Seoul Region)

Gemini API
,
1

  • Requested Models: gemini-3-pro-image-preview, gemini-3.1-flash-image-preview

  • GCP Project ID: giant project (Testing with 1 project first; more projects will follow later)

  • Used Region: asia-northeast3 (Seoul) / global

  • Endpoint Type: global

  • Billing Status: Enabled (Paid Tier Account)

  • Vertex AI API Status: Enabled

  • Service Account Role: Vertex AI User

1. Detailed Use Case & Purpose We are currently developing two generative AI-driven applications aimed at a future B2B distribution rollout:

  • Image Generation App: Utilizing Nano Banana Pro (powered by the Gemini 3 series image models) for high-fidelity image creation.

  • Automated Video Generation App: Utilizing Veo 3 to build an automated workflow for short-form/video content creation.

To ensure seamless production deployment for our enterprise clients, we require stable and uninterrupted access to the Gemini 3 preview models in our designated environment.

2. Estimated Resource Usage Our usage is projected across two distinct phases (Internal Beta and Production Launch):

  • Per-User Daily Allocation Estimates:

    • Veo 3: 20 requests per user / day

    • Nano Banana Pro (gemini-3.1-image-preview): 5 requests per user / day

  • Phase 1: Pre-deployment (Internal Testing & Beta)

    • Expected Users: ~20 concurrent users (Internal team & select beta testers)

    • Estimated Daily Volume:

      • Veo 3: ~400 requests / day

      • Nano Banana Pro: ~100 requests / day

  • Phase 2: Post-deployment (B2B Production Launch)

    • Expected Users: 80+ concurrent users (Enterprise clients)

    • Estimated Daily Volume:

      • Veo 3: 1,600+ requests / day

      • Nano Banana Pro: 400+ requests / day

Hi everyone,

We are experiencing a very confusing and frustrating issue while testing the Gemini 3 series models via Vertex AI, and we would appreciate some guidance.

The most challenging part is that the model works intermittently. Sometimes our requests go through successfully, but other times the system fails, disrupting our production workflow.

When it fails, we primarily encounter two types of issues:

  1. 403 Forbidden / Permission Denied errors.

  2. Quota / Resource Capacity Exceeded issues.

Our Environment & Context:

  • Model: Gemini 3 Series

  • Region: Seoul (asia-northeast3)

  • Account Status: We are using a fully enabled paid billing account (Paid Tier), so this shouldn’t be a simple free-tier restriction issue.

Since it works perfectly fine at times, we are confused as to why we suddenly hit 403 or quota errors without any changes to our IAM settings or code.

  • Is the Gemini 3 series in the Seoul region currently facing severe capacity constraints?

  • Does this model require a specific whitelist registration for asia-northeast3 even for paid users to guarantee stable access?

Any insights on how to resolve this instability or request proper allocation would be incredibly helpful.

Thank you!