Model is not found for API version v1beta — Veo and Imagen 404 in the Gemini API

Published 2026-09-23

HTTP 404 NOT_FOUND Do not retry as-is

The exact error

models/veo-2.0-generate-001 is not found for API version v1beta, or is not supported for predictLongRunning. Call ListModels to see the list of available models and their supported methods.

What it means

The Gemini API (generativelanguage.googleapis.com, version v1beta) answered 404 NOT_FOUND. The message covers two conditions at once: the model ID does not exist for this API, or it exists but not for the method you called. The method is named in the message, predictLongRunning for Veo and predict for Imagen. Public reports show at least three different causes behind the same text.

The model has been shut down on the Gemini API. This now covers every Imagen model. Google’s Imagen page says Imagen models are shut down and no longer available in the Gemini API, and points to Nano Banana instead. Earlier, the changelog recorded imagen-3.0-generate-002 as shut down on November 10, 2025, and in python-genai issue #967 an SDK contributor pointed a user who hit this 404 for that ID to that entry. For Veo, the deprecations page lists veo-2.0-generate-001, veo-3.0-generate-001 and veo-3.0-fast-generate-001 with June 30, 2026, and the Veo 3.0 preview IDs with November 12, 2025. Those are the earliest possible dates, and the page marks models already shut down in gray, so check it for any ID’s current status.
The model was never on the Gemini API. In python-genai issue #967, imagen-3.0-fast-generate-001 returned this 404 with the predict wording. The reporter had read Vertex AI’s model list. An SDK contributor replied that the list was for Vertex AI, while the reporter was using a Gemini API key.
The key’s project has no billing, in the Veo case. In google-gemini/cookbook issue #706, veo-2.0-generate-001 returned exactly the message at the top of this page. A repository collaborator asked whether the key was on the free tier, said in April 2025 that Veo was a paid-only feature, and reproduced the same 404 with a free key. The reporter confirmed no billing account was linked. So this 404 does not always mean the ID is wrong.

Merged variant: Invalid resource field value in the request. On Vertex AI (aiplatform.googleapis.com) a related failure comes back as 400 INVALID_ARGUMENT with reason RESOURCE_PROJECT_INVALID. In python-genai issue #2079, veo-3.1-fast-generate-001 failed this way through google-genai 1.64.0 with vertexai=True. The metadata named the v1beta1 PredictionService, while direct REST to v1 returned 200, so the reporter suspected SDK routing. Two responders could not reproduce it, an SDK contributor said v1 versus v1beta1 should not matter, and the issue closed as stale without a confirmed cause.

How to confirm it’s this

The status is 404 NOT_FOUND and the message names your model ID and a method.
Do what the message says and list the models your key can see. With the Python SDK, client.models.list() returns them. If your ID is missing, you are in case 1 or 2.
If the ID is listed but Veo still returns this 404, check whether billing is enabled on the Google Cloud project behind the API key (case 3).
Check which documentation you copied the ID from. IDs from cloud.google.com Vertex AI pages do not necessarily exist on the Gemini API.

Fix

Replace shut-down IDs with the replacement named on Google’s deprecations page. For Veo 3.0 and 2.0, the page points to the Veo 3.1 preview IDs on the Gemini API or the GA models on Gemini Enterprise Agent Platform (formerly Vertex AI).
Move Imagen calls to Nano Banana. Google’s Imagen page says to use gemini-2.5-flash-image or a Nano Banana 2 model such as gemini-3.1-flash-image, and to call generate_content instead of generate_images. The response shape changes too: images come back as content parts, not an image response object.
Use the backend the model belongs to. For a model documented only on Vertex AI, create the client with vertexai=True plus a project and location instead of an API key. Check Vertex AI’s own model list first.
Enable billing on the key’s project if you are calling Veo from a free-tier key, then check current pricing and rate limits in Google’s docs.
For the Vertex RESOURCE_PROJECT_INVALID variant, the SDK contributor in issue #2079 suggested pinning the API version to test the routing theory:

client = genai.Client(
    vertexai=True,
    project=project,
    location='us-central1',
    http_options={'api_version': 'v1'},
)

Retry or not

No. Nothing about the model name changes between attempts, so an identical retry gets the same 404. Google’s API errors page lists model_not_found with the advice to verify the model name or fall back to another model. That page documents the Interactions API’s lowercase codes, while predict and predictLongRunning return the gRPC-style NOT_FOUND shown here, but the advice carries over. An agent should treat this as a configuration error: fall back once to a model ID confirmed in the model list, then stop and report.

When it’s not this

A 403 when downloading a finished video is an authentication problem on the file URL, not a missing model.
A 401 saying API keys are not supported comes from calling an endpoint that needs OAuth.
A 429 with a limit of 0 means the model exists but your tier has no quota for it.
Location errors return 400 FAILED_PRECONDITION, not 404.

Error text last verified against the sources below on 2026-09-23.