Imagen 4 is shut down. Use Banana 2

Imagen 4 is shut down. It is not a live pick against GPT Image 2. Google's own Imagen page, last updated 17 September 2026, says Imagen is the legacy image model, it is now shut down, and it is no longer available in the Gemini API. The replacement is Nano Banana. On this catalog that means Nano Banana 2, not a leftover photoreal card. GPT Image 2 Text to Image is still the OpenAI instruction still. That is the live fork. An earlier draft of this URL treated Imagen 4 as a 4-credit 4K camera. That was catalog lag, not research.

This page is only the dead comparison. It is not Google's Great Model Cleanup (the June and August Gemini API cutoffs, including Veo 2 and 3). It is not GPT Image 2 vs Nano Banana 2 (the general stills pick). It is not Nano Banana 2 vs GPT Image 2 by job (rapid vs lettering vs edit). Stop opening Imagen 4 because a title named it.

Google shut the family, not a nickname

The dates are Google's. Read them as shutdowns, not as "maybe still in a picker."

On 15 June 2026 the Gemini API changelog deprecated imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, and imagen-4.0-fast-generate-001, with shutdown on 17 August 2026. Same entry as the Veo 2.0 and 3.0 cut. Imagen got nine weeks. Veo got two. Neither runway is still open.

On 17 August 2026 those three Gemini API endpoints hit the published shutdown date. Google's deprecations table (last updated 17 September 2026 UTC) still prints 17 August 2026 against all three ids, and names gemini-3.1-flash-image as the recommended replacement. Google footnotes that table dates are the earliest possible retirement. Plan against the changelog sentence anyway. A month later, the Imagen generate page no longer documents a live call. The first line is the product: Imagen models are shut down. Use Nano Banana.

Firebase's Imagen to Nano Banana migration guide, also last updated 17 September 2026, is louder than a catalog chip. All Imagen models shut down on 17 August 2026. The cutoff applies across Google, for both the Gemini Developer API and the Agent Platform Gemini API (formerly Vertex AI). The replacement table is the map, not a vibe:

• imagen-4.0-fast-generate-001 to gemini-3.1-flash-image with thinking level MINIMAL
• imagen-4.0-generate-001 to gemini-3.1-flash-image with thinking level HIGH
• imagen-4.0-ultra-generate-001 to gemini-3-pro-image

Vertex had an earlier clock. Google's Vertex notes put Imagen 3 and 4 generation endpoints behind a 30 June 2026 migration deadline after a 24 March 2026 deprecation. Firebase's model table also prints 30 June 2026 against the Imagen 4 ids. Two surfaces, two dates. Both dates are in the past. Do not pick Imagen 4 because a third-party catalog still prints 4 credits.

Imagen 4 launched as Google's photoreal stills family. The 20 May 2025 I/O post sold fabric, water, fur, and up to 2K. That post is history. A history page is allowed to say what the model was. A comparison page is not allowed to treat it as the Google camera you click this week.

The migration is not a rename

Teams lose a week here. They grep imagen-4.0-generate-001, paste gemini-3.1-flash-image, and ship. Google's Imagen page says that is not the job.

Three things change.

The id. Nano Banana 2 on Google's door is gemini-3.1-flash-image. The older Nano Banana line is gemini-2.5-flash-image, and that id itself carries a 2 October 2026 shutdown on the deprecations table. Do not migrate from a dead Imagen id onto a Banana id that is itself dated. On Versely the published Google still you should open is Nano Banana 2. The catalog scores it as Nano Banana 2 (Gemini 3.1 Flash Image).

The method. Imagen was client.models.generate_images. Banana is client.models.generate_content. Firebase's guide is even louder: you leave ImagenModel and you construct a GenerativeModel with a response modality of IMAGE.

The response. Imagen returned a generated-image object. Banana returns content parts that may contain image data. Tests that assert response.generated_images are the outage, not a style nit.

Config flags die with the family. Firebase's replacement table: addWatermark is gone (SynthID is always on for Gemini Image). numberOfImages is gone (one image per call; loop if you need a set). negativePrompt was already gone on Imagen 4. personGeneration is gone. imageFormat is gone (PNG). Aspect moves into an ImageConfig. That is why a "we swapped the string" deploy still fails in QA.

On this catalog, Banana 2 is not Imagen with a fruit name. Banana 2 is 4 credits discounted from 7, text-to-image plus image-to-image plus edit-image, up to 14 references, qualities 1K / 2K / 4K, no max_output_resolution field, billed flat. Imagen 4's old card was text-to-image only, 4 credits flat, photoreal, text-only, 4K listed, no edit path. You did not get a free edit lane by renaming the row. You got a different machine that happens to cost 4 credits on the discounted meter.

Nano Banana is the older 4-credit Google still: text-to-image only, 4K listed, no edit. Do not grab it because the name is shorter and the sticker matches. Rapid plus the option to keep editing is Banana 2.

The live fork is Banana 2 versus GPT Image 2

GPT Image 2 did not shut down. OpenAI launched ChatGPT Images 2.0 / gpt-image-2 on 21 April 2026. The published Versely card is GPT Image 2 Text to Image: 9 credits at SD/HD discounted (17 full), 17 at 4K discounted (33 full), high-fidelity, strong prompt adherence, up to 4K. That is still a live instruction still.

So the comparison people type ("Imagen 4 vs GPT Image 2") has a dead left side. Replace the left side. Then pick.

Pick Nano Banana 2 when you wanted the Google still, and you were about to click Imagen because the vendor is Google. Pick it when you will iterate. Banana 2 is billed as ultra-fast lightweight generation. The 4-credit sticker is a loop price, not a print-plate price. Pick it when you already have a file and the job is "change this, keep the object." That is an edit. Imagen 4 never had that category. Banana 2 does. Attach the reference. Do not re-roll a dead photoreal generate and hope the bottle stays. Pick it when you are not paying 17 credits to find out whether the crop works.

Pick GPT Image 2 Text to Image when the failure mode is ignored instructions: wrong object, wrong count, paraphrased headline, a layout you wrote in a paragraph. Pick it when the still has a quoted string in the frame you will proof at 100 percent. Pick it when you want the OpenAI adherence prior and you will pay 9 or 17 on purpose. Pick it when you are generating once, not sketching.

Pick Flux 2 Max when the job that used to send people to Imagen 4 was "this has to look photographed." Flux 2 Max is 6 credits flat, photorealistic, detailed, text-to-image only. It is not Google. It is the photoreal still that is still a live row. Banana 2 is the Google replacement. Flux is the camera replacement if photoreal materials were the whole brief.

Pick neither leftover Imagen comparison when the still is already a file. That is Banana 2 edit, or GPT Image 1.5 / GPT Image 1 Edit on the OpenAI side. Re-rolling a shut-down generate to move a necklace is how you spend a week on a model Google already turned off.

Open text to image and confirm the slug. nano-banana-2 or gpt-image-2-text-to-image. If the picker is still sitting on imagen-4 because an old post linked it, change the row. A listed slug is not a recommendation.

Credits are not USD. Do not convert 4, 6, 9, or 17 into a dollar figure. Google's old Gemini API post priced Imagen 4 Fast at $0.02 per image on Google's meter. That meter is gone. Banana 2 on this catalog is 4 credits discounted from 7. GPT Image 2 is 9 or 17. Those are Versely credits, 19 September 2026.

A still you already generated on Imagen is a file. Image-to-video does not care which vendor made frame one. It cares that frame one is locked. Keep the PNG if you would print it. Do not re-generate it on a dead row "to match the old look." If you need a new plate, start on Banana 2 or Flux 2 Max and proof it like a new still.

A catalog row is not Google availability

Versely still lists an imagen-4 card in models.json as of 19 September 2026: Google, text-to-image, 4 credits, 4K, photoreal, text-only. It is still in the published-models file. That is why the first draft of this URL treated it as a live camera. That was a research miss. Google's docs, Firebase's migration guide, and the deprecations table all say the family is shut down. A product catalog can lag a vendor shutdown. It cannot override one.

Treat a leftover Imagen 4 row the way you treat a leftover preview string: do not start a new job there. If a generate against that slug still returns a file, that is serving lag, not a stay of execution. Do not plan a campaign on it. Do not write a comparison that tells a stranger to pick it. Do not tell a client "Google's 4-credit 4K photoreal still." Google does not sell that API anymore.

The honest Google stills list on this catalog, 19 September 2026:

• Nano Banana 2: 4 credits discounted (7 full). Text-to-image plus image-to-image plus edit. Up to 14 refs. Rapid iterate. This is the replacement.
• Nano Banana: 4 credits. Text-to-image only. 4K listed. Older. No edit. Do not pick it because the name is shorter.
• imagen-4 and imagen-4-fast (2 credits, unpublished as a model page): leftover cards. Fast was never a published /models URL. Neither is a live Google pick.

Flow does not save you. Google Flow is a Veo surface, not a studio. Flow was also an Imagen surface. The Imagen surface shut down. You keep Veo 3.1 for video. You keep Banana 2 for stills. You do not keep living in a launcher because the landing page still says Imagen.

If you called Google directly, grep the org for imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001, and generate_images. Swap the id, the method, and the parser. If you generate here, skip the grep for those strings and pick Banana 2 or GPT Image 2 on the stills door. Then sign up and run the live row. Do not run a shutdown.

FAQ

Is Imagen 4 still a cheaper GPT Image 2?

No. It is not cheaper, because it is not a live Google generate you should start. GPT Image 2 T2I is 9 credits at SD/HD and 17 at 4K on the discounted matrix. Nano Banana 2 is 4 credits discounted. That is the Google sticker now. Four versus nine was never a quality ranking anyway. Photoreal materials and instruction following are different jobs. The dead row does not win either job.

Is Nano Banana 2 a drop-in for Imagen 4?

No. Google mapped the ids (imagen-4.0-generate-001 to gemini-3.1-flash-image with thinking HIGH). It did not map the product. Banana 2 edits, takes references, and is billed as a lightweight iterate. Imagen 4 was a text-only photoreal generate with no edit category. If the brief is "looks photographed, generate once," try Flux 2 Max at 6 credits, or generate on Banana 2 and proof like a camera. If the brief is "change this locked still," Banana 2 is the row Imagen never was.

Why does a catalog still show Imagen 4?

Because catalogs lag vendor shutdowns. Google's Imagen page (17 September 2026) says the model is shut down. The deprecations table still prints 17 August 2026 against the three Imagen 4 ids. Firebase says the shutdown applies across Google. A leftover 4-credit card is not a recommendation. Open Banana 2 or GPT Image 2. Leave the leftover row.