
Imagen 4 API Shut Down: How to Migrate, What It Costs
Imagen 4 API shut down on Vertex AI June 30, 2026 and in the Gemini API August 17. Google's deprecations table names Nano Banana 2.1. Migration paths and costs.
The Imagen 4 API is gone on both of Google's developer surfaces. Vertex AI discontinued imagen-4.0-generate-001 and its Fast and Ultra siblings on June 30, 2026, and the Gemini API followed on August 17. Google's deprecations table names gemini-nano-banana-2.1 as the replacement for all three, and on Arena's November 12 text-to-image board Nano Banana 2.1 sits 181 points above Imagen 4 Ultra.
TL;DR: Imagen 4 Standard, Ultra and Fast stopped answering on Vertex AI on June 30, 2026 and in the Gemini API on August 17, 2026. Google's recommended replacement in the Gemini API is
gemini-nano-banana-2.1, live here as the Nano Banana 2.1 API at $0.02 per 1K image. Our tier mapping: Fast to the Nano Banana 2 Lite API ($0.02), Standard to Nano Banana 2.1 at 1K or 2K, Ultra to Nano Banana 2.1 at 4K ($0.045). Even Lite outscores Imagen 4 Ultra on Arena by 103 points. If you followed Vertex's advice and moved to Gemini 2.5 Flash Image, note that the Gemini API gives that model an earliest shutdown of March 15, 2027.
Imagen 4 shutdown: Google's own pages still disagree
Read on November 16, 2026, three Google pages give three different answers about where Imagen 4 callers should go:
| Google page | What it says about Imagen 4 | Successor it names |
|---|---|---|
| Gemini API deprecations | Shut down August 17, 2026 (previews February 17, 2026) | gemini-nano-banana-2.1 |
| Imagen 4 model card | Still warns the endpoints "will be shut down" on August 17 | Gemini 3.1 Flash Image |
| Vertex AI Imagen 4 page | Discontinued June 30, 2026 | gemini-2.5-flash-image |
The deprecations table is the one to trust for the Gemini API, because it is the page Google keeps current: it moved its replacement column from Gemini 3.1 Flash Image to Nano Banana 2.1 after 2.1 launched on October 6, while the model card kept its August wording. The old ai.google.dev deprecations URL now redirects to the AI Studio docs.
Every Imagen 4 deprecated row in that table, and the Imagen 3 row with them, now carries a shutdown date in the past:
| Model ID | Released | Shut down (Gemini API) | Listed replacement |
|---|---|---|---|
imagen-4.0-generate-001 | June 24, 2025 | August 17, 2026 | gemini-nano-banana-2.1 |
imagen-4.0-ultra-generate-001 | June 24, 2025 | August 17, 2026 | gemini-nano-banana-2.1 |
imagen-4.0-fast-generate-001 | June 24, 2025 | August 17, 2026 | gemini-nano-banana-2.1 |
imagen-4.0-generate-preview-06-06 | June 24, 2025 | February 17, 2026 | gemini-nano-banana-2.1 |
imagen-4.0-ultra-generate-preview-06-06 | June 24, 2025 | February 17, 2026 | gemini-nano-banana-2.1 |
imagen-3.0-generate-002 | February 6, 2025 | November 10, 2025 | gemini-nano-banana-2.1 |
Google defines a shut-down model as completely turned off, with the endpoint gone, so no standalone Imagen model answers through the Imagen API in AI Studio anymore.
Vertex AI shut Imagen 4 first, and its successor has a date too
Enterprise teams on Vertex AI hit the wall seven weeks earlier. Google's Vertex page for Imagen 4 lists imagen-4.0-generate-001 as released there on August 14, 2025 and discontinued on June 30, 2026, with the Fast and Ultra endpoints, every Imagen 3 endpoint and the older imagegeneration@002 to @006 builds pointed at gemini-2.5-flash-image.
That recommendation has a catch on the Gemini API side. The same deprecations table lists gemini-2.5-flash-image, the original Nano Banana, with an earliest shutdown of March 15, 2027 and gemini-3.1-flash-lite-image (Nano Banana 2 Lite) as its replacement. Google says those dates are the earliest possible ones and that it gives notice before the exact day. Still, a team that moved from Imagen 4 to 2.5 Flash Image in June is on a model with a scheduled end, and will likely migrate twice.
On UnificAlly the original Nano Banana still runs as google/nano-banana, and its catalog entry already points to Nano Banana 2.1 as the replacement. New work should not start there.
Is Gemini 3.1 Flash Image deprecated?
Not by Google's own table. gemini-3.1-flash-image, which UnificAlly serves as the Nano Banana 2 API (google/nano-banana-2), sits in the active Nano Banana table with "No shutdown date announced" and gemini-nano-banana-2.1 as the recommended replacement. It is not in the deprecated-models sub-table where Imagen and the old previews live. Teams that migrated from Imagen 4 in August, when Gemini 3.1 Flash Image was the named target, keep working with no forced move. The gemini-nano-banana-2.1 migration from there is a one-line model change on our API, because the request fields match.
Imagen 4 vs Nano Banana on Arena: even Lite wins
Imagen 4 was still on Arena's text-to-image board when we pulled it on November 16 (board dated November 12, 2026, 6,934,037 votes):
| Model | Arena rank | Score | Status |
|---|---|---|---|
| Nano Banana 2.1 | #7 | 1327 (preliminary) | Live on UnificAlly |
| Nano Banana 2 (Gemini 3.1 Flash Image) | #13 | 1256 | Live on UnificAlly |
| Nano Banana 2 Lite | #16 | 1249 | Live on UnificAlly |
| Nano Banana Pro (2K) | #18 | 1246 | Live on UnificAlly |
| Nano Banana (Gemini 2.5 Flash Image) | #38 | 1148 | Live, replacement set |
| Imagen 4 Ultra | #40 | 1146 | Shut down |
| Imagen 4 | #47 | 1128 | Shut down |
| Imagen 3 | #69 | 1057 | Shut down |
On the November 12 Arena text-to-image board, Nano Banana 2 Lite scores 1249 against 1146 for Imagen 4 Ultra, so the 1K-only Lite model beats the top Imagen tier by 103 points. Imagen 4 Fast never had its own row. The 2.1 score still rests on fewer than 5,000 votes, but its lead over Ultra (181 points) is many times the ±9 margin.
Pro is the one result that changed our mapping. In August we sent Imagen 4 Ultra traffic to Pro, the family's quality tier. On today's board Pro trails 2.1 by 81 points while costing $0.07 at 1K against $0.02 for 2.1, so Ultra traffic goes to 2.1 at 4K. Our Nano Banana 2.1 vs Nano Banana 2 comparison covers the two current models in depth, and the GPT Image 2 vs Nano Banana comparison covers the models ranked above them.
Imagen 4 replacement by tier, with prices
UnificAlly never served Imagen 4, so nothing broke here in June or August; these are successors, not the same endpoints:
| If you were calling | Our pick | Price per image |
|---|---|---|
imagen-4.0-fast-generate-001 | Nano Banana 2 Lite, 1K only | $0.02 |
imagen-4.0-generate-001 | Nano Banana 2.1 at 1K or 2K | $0.02 / $0.03 |
imagen-4.0-ultra-generate-001 | Nano Banana 2.1 at 4K | $0.045 |
Lite keeps Fast's job, but only at 1K: its docs allow no other resolution. Google lists gemini-3.1-flash-lite-image with no shutdown date and makes it the named successor for the original Nano Banana, so it is a stable target. If your Fast pipeline produced larger files, use 2.1 at 2K with thinking_level set to low instead (next section).

What changes in an Imagen 4 migration request
Imagen 4 took a text prompt of up to 480 tokens, accepted no input images, and returned 1 to 4 images per request, according to Google's model card. Nano Banana 2.1 on UnificAlly flips most of that:
| Imagen 4 (Gemini API) | Nano Banana 2.1 (UnificAlly) |
|---|---|
| Text in only | prompt plus image_urls, up to 10 reference images |
| 1 to 4 images per request | One image per task; the docs list no count field, so send four tasks for four images |
| Three model IDs for speed and quality | One model ID; resolution 1K, 2K or 4K picks the size |
| Fast tier for speed | thinking_level low, about twice as fast as the default high, at the same price |
| Aspect ratio per request | aspect_ratio 1:1, 16:9, 9:16, 4:3, 3:4 or auto (follows the first input image) |
The thinking_level row is the useful surprise. The model plans the image first, the bill is identical at all three settings, and high is the setting the docs recommend for exact text, counts and labelled diagrams, the work Imagen 4 was marketed for. Drop to low for drafts and simple product shots.
A Standard-tier replacement call, with the Ultra variant one field away:
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "google/nano-banana-2.1",
"input": {
"prompt": "A product hero shot of a ceramic mug on a walnut desk, soft morning light",
"aspect_ratio": "16:9",
"resolution": "1K",
"thinking_level": "medium"
}
}'
You get a task_id back straight away; read the finished image from GET /v1/tasks/<task_id> or have it posted to a callback_url. For the Ultra mapping, send "resolution": "4K".
Imagen 4 alternative costs at catalog scale
Because one Imagen 4 request could return four images, the honest comparison is per image, and a four-image request becomes four tasks: $0.08 on Nano Banana 2.1 at 1K, $0.18 at 4K, or $0.08 on Lite.
Regenerating a 1,000-image catalog costs $20 on Lite, $20 on Nano Banana 2.1 at 1K and $45 at 4K. Rendering everything at 1K and only the hero shots at 4K keeps most of the bill at the 1K rate, since resolution is a field on the same model. You pay only for images that finish, and a new account's $0.50 signup balance covers 25 Nano Banana 2.1 images at 1K, enough to run a test set of your old Imagen prompts first.
Why Google folded image generation into Gemini
Imagen was a standalone diffusion line driven by a prompt and a few parameters. The Gemini image models reason over the prompt, accept reference images and edit in the same call, and Google has been retiring the standalone stack for over a year: Imagen 3 in November 2025, the Imagen 4 previews in February 2026, the GA endpoints this summer.
Video is going the same way. Google's Veo 3.1 preview IDs in the Gemini API shut down on October 22, 2026 with gemini-omni-1.2-flash named as the replacement, and gemini-omni-1.1-flash starts routing to 1.2 on December 3, 2026. Our Gemini Omni 1.1 Flash post covers that line. Betting new image work on a Gemini-family ID is the choice least likely to need a third migration.
Imagen 4 FAQ
Is the Imagen 4 API still available?
No. Vertex AI discontinued Imagen 4 Standard, Fast and Ultra on June 30, 2026, and the Gemini API shut down the same three endpoints on August 17, 2026, after retiring the preview builds in February. Any request naming an imagen-4.0 model ID now fails.
What replaced Imagen 4?
In the Gemini API, gemini-nano-banana-2.1, the replacement Google's deprecations table lists for every Imagen row. Vertex AI's page names gemini-2.5-flash-image instead, and Google's Imagen 4 model card still says Gemini 3.1 Flash Image. On UnificAlly, Nano Banana 2.1 costs $0.02 per 1K image.
I already migrated to Nano Banana 2. Do I need to move again?
No deadline forces it. Google's table shows no shutdown date for gemini-3.1-flash-image and keeps it out of the deprecated-models list, and UnificAlly still serves it as google/nano-banana-2. Switching to 2.1 is worth doing on your next release: it ranks 71 points higher on Arena and takes the same request fields.
Is Nano Banana the same as Imagen?
No. Imagen was Google's standalone diffusion family; Nano Banana is image generation inside Gemini models, with reasoning before drawing and reference images. Nano Banana 2.1 is the model Google's deprecations table now names for Imagen callers, which makes it the successor in practice rather than a rename.
Do Imagen 4 prompts work on the replacement?
Mostly, but test before you switch production traffic. Imagen capped prompts at 480 tokens and took no images, while Nano Banana 2.1 is prompt-first with up to 10 references and a thinking_level setting, so the same words can land differently. Ten representative prompts at high usually show the gaps.
Can I still use Imagen 4 anywhere?
Not from Google. Both of Google's developer surfaces have switched it off: Vertex AI on June 30, 2026 and the Gemini API on August 17, 2026. Imagen weights were never released, so self-hosting was never possible either.
How much does the replacement cost per image?
On UnificAlly: $0.02 on Nano Banana 2 Lite, and $0.02, $0.03 or $0.045 on Nano Banana 2.1 at 1K, 2K or 4K. Nano Banana Pro is $0.07 at 1K. Every thinking_level costs the same, and failed tasks are not billed.
The models covered here are live as the Nano Banana 2.1 API, the Nano Banana 2 Lite API and the Nano Banana Pro API, each with a playground for your old Imagen prompts.


