Vertex AI Gemini 3.8 Flash generateContent returns 403 PERMISSION_DENIED after previous HTTP 200 responses

Hello,

I’m seeing a new HTTP 403 / PERMISSION_DENIED response when calling Vertex AI generateContent with Gemini 3.8 Flash.

Configuration:

  • Project: (PII Removed by Staff)

  • Service account: (PII Removed by Staff)

  • Model: gemini-3.8-flash

  • Location: eu

  • Endpoint: aiplatform.eu.rep.googleapis.com

  • Operation: generateContent

A known failed pilot response was received at:

2026-11-02T09:16:53.850613Z

The same project, service account, model, location, and endpoint previously returned HTTP 200 successfully.

What I have already verified:

  • aiplatform.googleapis.com is enabled.

  • Billing is enabled.

  • Service-account impersonation succeeds.

  • The service account has roles/aiplatform.user.

  • The IAM binding has no condition.

  • No project-level IAM deny policy was found.

  • The project has no parent folder or organization in the observed hierarchy.

  • Narrow Cloud Logging searches around the failure returned no visible matching audit entry.

  • A separate one-shot synthetic request also reproduced HTTP 403 / PERMISSION_DENIED.

The synthetic test contained no customer or sensitive data.

Could someone help identify what provider-side check might be causing this?

In particular:

  1. Is there any additional access, entitlement, eligibility, abuse-prevention, or project/account restriction for gemini-3.8-flash on Vertex AI?

  2. Can PERMISSION_DENIED occur even when roles/aiplatform.user and aiplatform.endpoints.predict are available?

  3. Have access requirements for Gemini 3.8 Flash changed recently?

  4. Is there another read-only diagnostic that can reveal the denied permission or internal policy check without enabling additional APIs?

I can provide sanitized timestamps and local evidence if needed, but I will not post access tokens, Authorization headers, credentials, or customer data.

Thank you.