[BUG] Paid Tier 1 — gemini-3-pro-image persistent HTTP 5xx across Interactions and GenerateContent

Gemini API
, , ,
1

Hello Gemini API team,

I am experiencing persistent server-side failures with gemini-3-pro-image on a correctly configured Paid Tier 1 / Prepay Google AI Studio project.

I have performed extensive troubleshooting and have now reproduced the problem across both officially supported generation surfaces.

Environment

  • Google AI Studio project: MERKOA Content Pilot
  • Project ID: gen-lang-client-01787670..
  • Billing: Paid Tier 1 / Prepay
  • Available prepaid credit: approximately USD 9.99
  • API key is attached to this exact project
  • Python SDK: google-genai==2.19.0
  • Model: gemini-3-pro-image / Nano Banana Pro
  • Request: text-to-image, 2K, 1:1

Account / quota checks already completed

  • Billing is active and there are no billing warnings.
  • Prepaid balance is positive.
  • The API key belongs to the affected paid project.
  • Rate limits are nowhere near exhaustion.
  • For Nano Banana Pro, AI Studio currently shows approximately:
    • RPM peak: 6 / 20
    • TPM peak: 3.14K / 100K
    • RPD peak: 19 / 250
  • Account information / age verification is complete.
  • Google AI Studio shows the project as Paid Tier 1 / Prepay and in normal state.

Model discovery works

Real non-generative metadata requests succeeded.

models.list() successfully returned:

  • gemini-3-pro-image
  • gemini-3-pro-image-preview
  • nano-banana-pro-preview

A direct models.get() for the stable gemini-3-pro-image target also succeeded.

Therefore:

  • the API key is valid;
  • the project can reach the Gemini API;
  • the stable model is visible to this account;
  • model metadata access is working.

Generation failures

I have performed three controlled real generation attempts. No automatic application retries or fallback models were used.

Attempt 1 — Interactions API



client.interactions.create(
    model="gemini-3-pro-image",
    input=prompt,
    response_format={
        "type": "image",
        "aspect_ratio": "1:1",
        "image_size": "2K",
    },
)

Result:

google.genai._gaos.lib.compat_errors.InternalServerError

HTTP class: 5xx.

No image was returned.

During this first attempt we later discovered that the Interactions transport had SDK-level automatic retries enabled by default, so the exact number of underlying HTTP attempts from this historical run could not be recovered.

Attempt 2 — Interactions API after explicitly disabling retries

We corrected the transport configuration to:

HttpRetryOptions(attempts=1)

This guarantees:

  • one application invocation;
  • maximum one underlying HTTP attempt;
  • zero automatic transport retries.

The exact same stable model and request configuration were used.

Result:

google.genai._gaos.lib.compat_errors.InternalServerError

Again, a real HTTP 5xx server-side error.

The call completed/faulted in approximately 38 seconds.

No image was returned.

Attempt 3 — GenerateContent API

To determine whether this was specific to the Interactions API, I tested the same stable model through the classic GenerateContent surface:



client.models.generate_content(
    model="gemini-3-pro-image",
    contents=prompt,
    config=types.GenerateContentConfig(
        response_modalities=["IMAGE"],
        image_config=types.ImageConfig(
            aspect_ratio="1:1",
            image_size="2K",
        ),
    ),
)

Retry configuration was again explicitly limited to one total HTTP attempt.

Result:

google.genai.errors.ServerError

Failure occurred after approximately 3.03 seconds.

No image was returned.

Important observations

The failure therefore reproduces:

  • with the stable gemini-3-pro-image model;
  • through Interactions API;
  • through GenerateContent API;
  • with automatic retries disabled;
  • while metadata/list/get requests work normally;
  • with valid Paid Tier 1 billing and positive prepaid credit;
  • while well below RPM/TPM/RPD limits.

No successful image generation has occurred, and the failed requests have not produced any significant corresponding usage charge visible in AI Studio.

What I am asking Google to check

Could someone from the Gemini API team please investigate whether there is a project/account-specific backend issue affecting image generation for gemini-3-pro-image?

Specifically, could you check:

  • whether this project has any hidden model-generation restriction or backend state not visible in AI Studio;
  • whether there is a known issue affecting Paid Tier 1 / Prepay projects with Gemini Pro Image;
  • whether gemini-3-pro-image is currently expected to work for this project through both Interactions and GenerateContent;
  • whether additional project-side activation, allowlisting, or entitlement is required;
  • whether you need timestamps, local trace IDs, project number, or additional sanitized diagnostics from me.

I can provide additional identifiers privately if a Google team member needs them.

For security reasons I will not publish my API key or billing account ID.

Thank you.

2

Hello @Merkoa ,

To help isolate the issue, please test image generation using the default 1K resolution (omit the image_size parameter or set it to “1K”).
This will help confirm whether the error is specific to the 2K resolution setting on your project.
You can verify supported configurations in the Image Generation Guide at https://ai.google.dev/gemini-api/docs/image-generation.
Please let us know if 1K requests succeed or produce the same error.