Imagen 3 API: Shutdown Status and What to Use Instead

If you found an Imagen 3 API tutorial and the code fails, the code isn't broken. The model is gone. On the Gemini API, imagen-3.0-generate-002 was shut down on November 10, 2025. Google then deprecated every remaining Imagen model, including the Imagen 4 models that replaced it, with a shutdown date of August 17, 2026 that applies across Google. The path Google now points to is its Gemini image models, the ones it brands "Nano Banana."

Here's the short version, checked against Google's own docs on September 17, 2026:

• All Imagen models: Google's migration guide says they are deprecated and shut down as early as August 17, 2026, on both the Gemini Developer API and the Agent Platform Gemini API (formerly Vertex AI). That date has passed.
• Imagen 3 on the Gemini API: imagen-3.0-generate-002 was already shut down on November 10, 2025.
• What to use instead: gemini-3.1-flash-image (Nano Banana 2) for most jobs, or gemini-3-pro-image (Nano Banana Pro) for the heaviest ones.

The rest of this post covers the model IDs you'll see in old code, what changes in the API when you move, and a checklist for the move itself.

Imagen 3 Status by Surface

Imagen reached developers through three main surfaces. The shutdown covers all of them. Google Cloud's docs now also call Vertex AI the Gemini Enterprise Agent Platform, and Firebase calls it the Agent Platform Gemini API, so you may see any of those names.

The cross-surface date comes from Firebase AI Logic's Imagen migration guide. It says all Imagen models are deprecated and shut down as early as August 17, 2026, and that this "is applicable across Google and for both the Gemini Developer API and Agent Platform Gemini API (formerly Vertex AI)."

The Gemini API deprecations page agrees. It marks imagen-3.0-generate-002 as already shut down and lists August 17, 2026 for all three Imagen 4 models, and the Imagen page in the Gemini API docs tells you to move to Nano Banana. On Vertex AI, the Imagen 3 model page carries a deprecation notice. That notice covers imagen-3.0-generate-001, imagen-3.0-fast-generate-001, imagen-3.0-generate-002, both imagen-3.0-capability versions, and the three Imagen 4 GA models. Whichever surface you're on, treat every Imagen endpoint as retired.

What Imagen 3 Was, and the Model IDs in Old Code

Google announced Imagen 3 at I/O on May 14, 2024, calling it its highest quality text-to-image model at the time. It first appeared as a private preview in ImageFX. It reached Vertex AI for approved users in July 2024 and opened to all Vertex users in December 2024. On February 6, 2025, Imagen 3 arrived in the Gemini API as imagen-3.0-generate-002.

That timeline is why tutorials from 2024 and 2025 use so many different IDs. Here's what you're likely to find:

You can also spot Imagen code by its call shape, even when the model name sits in a config file:

• Python client.models.generate_images(...) or JavaScript ai.models.generateImages(...)
• REST calls ending in :predict, with an instances array and a sampleCount parameter
• Firebase imagenModel(...) or getImagenModel(...)
• A REST host of generativelanguage.googleapis.com (Gemini API) or LOCATION-aiplatform.googleapis.com (Vertex AI)

What error to expect

Google doesn't publish an exact error message for calling a retired Imagen model. It does document the behavior. The Gemini API deprecations page says a shut-down model "is completely turned off, and the endpoint is no longer available." The Firebase AI Logic models page is more specific: a request using a retired model name returns a 404 error. If your old script suddenly gets a not-found response on a call that used to work, a retired model ID is the first thing to check.

Two tempting fixes that don't last

Swapping in Imagen 4. The Gemini API changelog entry for June 15, 2026 announced that imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, and imagen-4.0-fast-generate-001 would shut down on August 17, 2026, and Google's migration guide applies that date to every Imagen model on every surface. Imagen 4 isn't a destination any more.

Swapping in gemini-2.5-flash-image. Some Google pages still name it. Vertex AI's Imagen deprecation notice does, and so does the migration note on the Gemini API Imagen page. It has its own clock, though. The Gemini API deprecations page lists its shutdown as October 2, 2026. The Vertex AI model versions page lists its retirement as March 15, 2027. Go straight to a Gemini 3.x image model instead.

The Replacement: Gemini Image Models ("Nano Banana")

"Nano Banana" is Google's name for the image generation built into Gemini. The Gemini API image generation guide lists four models:

For an imagen-3.0-generate-002 call, follow Google's replacement chain. The Gemini API deprecations table points Imagen 3 to imagen-4.0-generate-001, and points that model to gemini-3.1-flash-image. On the Gemini API, gemini-3.1-flash-image and gemini-3-pro-image went GA on May 28, 2026, and the deprecations page lists no shutdown date for either. On Vertex AI, the model versions page lists both as available until May 28, 2027 or later.

What Changes When You Migrate

This isn't a model-name swap. Imagen was a dedicated image endpoint. Nano Banana is a multimodal Gemini model that happens to return images, so the request, the parameters, and the response all change.

Call shape: generateImages becomes an Interactions or generateContent call

On the Gemini API, the image generation guide now uses the Interactions API. Google's Interactions API overview says the original generateContent API is "now considered legacy" but "remains fully supported." Firebase AI Logic's migration samples use generateContent with the response modality set to IMAGE.

Here's what old Imagen 3 code on the Gemini API usually looks like, following the shape of Google's Imagen samples:

from google import genai
from google.genai import types

client = genai.Client()

response = client.models.generate_images(
    model="imagen-3.0-generate-002",
    prompt="Robot holding a red skateboard",
    config=types.GenerateImagesConfig(number_of_images=4),
)

And here's the replacement, built from the samples in Google's current image generation guide:

from google import genai
import base64

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-3.1-flash-image",
    input="Robot holding a red skateboard",
    response_format={"type": "image", "aspect_ratio": "16:9", "image_size": "2K"},
)

with open("generated_image.png", "wb") as f:
    f.write(base64.b64decode(interaction.output_image.data))

Where Imagen returned a list of generated images, the Interactions API gives you output_image, which Google's guide describes as the last generated image block.

Parameters: what maps and what disappears

Firebase AI Logic's Imagen migration guide has the clearest parameter mapping. The Gemini API guide fills in the sizes and ratios.

Two more differences can catch you out. The Gemini API's Imagen docs list English-only prompts, while the Nano Banana guide lists more than a dozen languages that perform best. And Gemini 3 image models are thinking models. Thinking can't be turned off in the API, and Google's guide says thinking tokens are billed by default. You can set thinking_level to minimal or high.

SynthID watermarking stays

Nothing changes on provenance. Google's docs say every image from Imagen and every image from the Nano Banana models carries a SynthID watermark. SynthID is invisible, so it won't change how your images look, and you can't switch it off on the new models either.

Imagen 3 Migration Checklist

1. Find every Imagen call. Search your code for imagen-3.0, imagen-4.0, imagegeneration@, generate_images, generateImages, :predict, and imagenModel. Check config files, environment variables, and remote config too.
2. Identify the surface. Your API host or SDK tells you whether a call runs through the Gemini API, Vertex AI, or Firebase AI Logic. The shutdown covers all three, but the SDK you rewrite against differs.
3. Pick a 3.x target. Use gemini-3.1-flash-image for most Imagen 3 and Imagen 4 Standard or Fast work. Use gemini-3-pro-image where you used Imagen 4 Ultra. Skip gemini-2.5-flash-image.
4. Rewrite the call. Move from generate_images to the Interactions API or generateContent with image output.
5. Replace numberOfImages with a loop. Each image is now its own request, so check your rate limits and cost assumptions.
6. Move ratio and size settings. Put them in response_format or ImageConfig. If you use Firebase, ImageConfig needs SDK versions from early May 2026 or later.
7. Delete dead parameters. Remove negativePrompt, addWatermark, and personGeneration. Firebase apps also need to move from ImagenSafetySettings to SafetySetting.
8. Rewrite response handling. Read image data from the response parts or output_image. Handle the case where no image comes back, which Firebase exposes as a NO_IMAGE finish reason.
9. Re-test your prompts. A multimodal model reads prompts differently than Imagen did. If your prompts relied on Imagen's quirks, rework them. Our guide to writing prompts for AI images covers the basics, and if you're rebuilding prompts from reference images, see image-to-prompt tools.
10. Make the model name easy to change. Firebase recommends Remote Config or server prompt templates for exactly this reason. Whatever you use, don't hard-code a model ID into a shipped client.
11. Watch the source pages. The deprecations page and changelog are where Google posts the next shutdown.

If image generation is one step in a larger Gemini build, our overview of building agents on Gemini covers the rest of Google's stack. And if you want a text model to trigger image generation, the LLM tool calling explainer shows how a model asks your app to run a function like this one.

Old tutorials break for a reason nobody warns you about: the setup worked when someone wrote it down, then the model underneath was retired. If you'd rather start from an AI setup someone already got running than rebuild one from a dated guide, Taku lets you mirror a working AI app or workflow into your desktop workspace, run it, and keep it. Taku is in Beta, and the Mac app is available now.

Key points

• imagen-3.0-generate-002 was shut down on the Gemini API on November 10, 2025.
• Google's migration guide says all Imagen models, Imagen 3 and Imagen 4 included, shut down as early as August 17, 2026, across Google, on both the Gemini Developer API and the Agent Platform Gemini API (formerly Vertex AI).
• The replacement is Google's Gemini image models: gemini-3.1-flash-image (Nano Banana 2) or gemini-3-pro-image (Nano Banana Pro).
• Don't migrate to Imagen 4 or gemini-2.5-flash-image. Imagen 4 falls under the same shutdown, and gemini-2.5-flash-image has its own retirement dates.
• The API changes shape: no numberOfImages, no negative prompts, aspect ratio moves into the response config, and every image still carries a SynthID watermark.

FAQ

Is the Imagen 3 API still available?

No. Google's deprecations page shows imagen-3.0-generate-002 shut down on the Gemini API on November 10, 2025. Google's migration guide says all Imagen models shut down as early as August 17, 2026, on both the Gemini Developer API and Vertex AI.

What replaced Imagen 3 in the Gemini API?

Google first named Imagen 4 (imagen-4.0-generate-001), but Imagen 4 is now deprecated too. Its listed replacement is gemini-3.1-flash-image, Nano Banana 2. For work that needed Imagen 4 Ultra, Google's Firebase guide maps to gemini-3-pro-image.

Can I still use Imagen 3 on Vertex AI?

No. Google's migration guide says the August 17, 2026 shutdown of all Imagen models applies to the Agent Platform Gemini API (formerly Vertex AI) as well as the Gemini Developer API. Move to a Gemini 3.x image model, and watch the Vertex AI release notes for later changes.

How do I generate multiple images with Nano Banana?

Make one request per image. Firebase's migration guide says Gemini image models return a single image and don't support numberOfImages. It also warns that candidate count doesn't work as a substitute. The Gemini API guide adds that the model won't always follow the exact number of images a prompt asks for.

Do Nano Banana images have a SynthID watermark like Imagen 3?

Yes. Google's image generation guide says all generated images include a SynthID watermark. Firebase's guide says you can't turn it off, so the old addWatermark parameter has no equivalent.

What is the model ID for Nano Banana 2?

gemini-3.1-flash-image. Nano Banana Pro is gemini-3-pro-image, and Nano Banana 2 Lite is gemini-3.1-flash-lite-image. The preview IDs gemini-3.1-flash-image-preview and gemini-3-pro-image-preview have a listed shutdown date of June 25, 2026, so update any code that still uses them.