What happened to Google Imagen on 17 August 2026?
All Imagen models are deprecated and shut down as early as 17 August 2026, across both the Gemini Developer API and the Agent Platform Gemini API (formerly Vertex AI). Google's replacement is the Gemini 3.x Image family, marketed as "Nano Banana". Anything still calling an Imagen endpoint stops returning images.
If an image feature you rely on started failing this week, check the model name before you rewrite the prompt.
You do not need to be a developer to be affected. Plenty of the automations practitioners run on Zapier, Make and n8n, plus a long tail of small SaaS tools and Figma plugins, quietly call an Imagen endpoint under the hood. When the endpoint dies, the tool just stops producing images and often shows a generic error.
Google published the official replacement map in its Imagen to Nano Banana migration guide, last updated 13 August 2026:
--- imagen-4.0-fast-generate-001 becomes gemini-3.1-flash-image at thinking level MINIMAL
--- imagen-4.0-generate-001 becomes gemini-3.1-flash-image at thinking level HIGH
--- imagen-4.0-ultra-generate-001 becomes gemini-3-pro-image
--- imagen-3.0-capability-001 becomes gemini-3.1-flash-image
The same page carries a second warning worth diarising now: Gemini 2.5 models shut down in October 2026. If your workflow still names a 2.5 model anywhere, you have roughly two months.
What is Nano Banana, and how is it different from Imagen?
Nano Banana is Google's brand name for its Gemini 3.x image models. Unlike Imagen, which was a dedicated image generator with its own API surface, Nano Banana is a normal Gemini model that happens to return an image. You send it a prompt through the same content-generation call you would use for text, with the response modality set to image.
That architectural change is the reason so many settings disappeared. Imagen had image-specific parameters. Gemini has one generic generation config with a small image config nested inside it.
In practice the swap is an upgrade for editing and a downgrade for batch control. Independent 2026 model roundups consistently rate Nano Banana highest on edit fidelity and inpainting, meaning it changes the one thing you asked it to change and leaves the rest of the frame alone. That is exactly the behaviour Imagen was weakest at.
What you lose is covered in the next section, and it is the part most people discover at the worst possible moment.
Which five Imagen features does Nano Banana not have?
Five Imagen parameters have no Nano Banana equivalent: negative prompts, image count, image format, watermark control, and person-generation control. Google documents each as "not supported" rather than renamed. If your prompt template relies on any of them, the template needs rewriting, not remapping.
1. negativePrompt is gone. You can no longer list what you do not want. Google notes negative prompts have been unsupported since imagen-3.0-generate-002 anyway, so many people were already passing a parameter that did nothing.
2. numberOfImages is gone. Nano Banana always returns a single image. If you were generating four variations per call and picking the best, you now loop four calls. Google explicitly says candidate count is not a substitute.
3. imageFormat is gone. Output is always PNG. If your pipeline expected JPEG or WebP, add a conversion step, and expect your storage bill and page weight to rise until you do.
4. addWatermark is gone. Every Nano Banana image carries a SynthID watermark, and you cannot turn it off. For most marketing work this is invisible and irrelevant. For anyone whose client contract forbids embedded provenance data, it is a genuine blocker worth raising early.
5. personGeneration is gone. Nano Banana allows images of people by default, with no dial to restrict it. Teams that used that setting as a compliance guardrail need a review step instead of a parameter.
How do you rewrite a negative prompt so it still works?
Convert every exclusion into a positive description of what should be there instead. "No text" becomes "clean untouched surface". "Not cartoonish" becomes "photographic, natural skin texture, real camera optics". Image models weight what you describe, so naming an unwanted object can summon it. Describing its replacement removes it.
This is the single highest-value habit to build this week, because it is the fix for the most common Nano Banana complaint, which is that images look close to the Imagen output but not the same.
Here is a complete template you can paste and adapt. Fill the five bracketed slots and delete nothing else.
Try this prompt:
Generate a single 16:9 image.
SUBJECT: [what is in frame, one sentence, concrete nouns only]
SETTING: [where it is, including surface, background and depth]
LIGHT: [direction, quality and colour temperature, for example "soft window light from camera left, cool daylight"]
CAMERA: [lens and distance, for example "50mm, waist-up, shallow depth of field"]
SURFACES: [what every area of the frame is made of, especially any area you previously banned things from, for example "plain matte wall, no signage, uninterrupted"]
Render photographically with natural imperfections. Every surface listed above must be present exactly as described. Do not add objects, people, text or logos that are not named above.
The SURFACES line is doing the work your old negative prompt used to do. Instead of "no logos on the wall", you assert that the wall is plain, matte and uninterrupted, which leaves the model nowhere to put a logo.
The final sentence is a closed-world instruction. It is not magic, and it will not hold on every generation, but across a batch it measurably reduces the stray coffee cups and invented signage that appear when a prompt leaves parts of the frame undescribed.
Which AI image model should you use for which job in 2026?
There is no single best image model in 2026. Independent comparison roundups published this year converge on a routing answer rather than a winner: match the model to the job, because each one leads on a different axis and the gap between them is now larger than the gap between prompt styles.
The rough consensus across 2026 model roundups looks like this:
--- Nano Banana (Gemini 3.x Image) for editing an image you already have. Best-in-class inpainting and edit fidelity, and now the default if you were on Imagen.
--- Midjourney V8.1 for art direction and cinematic concepts. Still the aesthetic leader, still the weakest at doing exactly what you asked.
--- FLUX.2 for photorealism and product shots. Strong colour accuracy, multi-reference composition and believable lighting.
--- GPT Image 2 as the general all-rounder for production and editing when you want one tool rather than four.
--- Ideogram 4 when the image contains readable text.
--- Recraft V4.1 for brand graphics, icons and anything that has to sit inside a design system.
The pattern worth stealing from teams running paid social at volume is that they chain models rather than choose one. A common pipeline is Midjourney for the initial art direction, then Nano Banana to edit the winning frame into the four variants the campaign actually needs. Neither model does both halves well.
If you have been loyal to one generator for two years, this is the assumption to re-test. The routing decision is now worth more than prompt-craft inside any single tool.
What does AI image generation actually cost per image?
Cost depends far more on how you access a model than on which model it is. The same Nano Banana Pro image can cost roughly thirteen cents through the pay-as-you-go API or well under a cent through a subscription, purely because subscriptions bundle a daily allowance you may or may not use.
Figures reported by third-party pricing trackers in 2026, useful as a directional guide rather than a quote:
--- Nano Banana Pro via official API: around US$0.134 per 1K or 2K image, and around US$0.24 per 4K image
--- Google AI Plus: around US$7.99 per month
--- Google AI Pro: around US$19.99 per month, reported at roughly 100 images per day
--- Google AI Ultra: around US$99.99 per month
--- Midjourney Basic: US$10 per month for roughly 3.3 GPU hours
--- Midjourney Standard: US$30 per month for roughly 15 GPU hours
--- Midjourney Pro: US$60 per month for roughly 30 GPU hours
The practical test is your own volume. Divide your monthly image count into the subscription price and compare it against the per-image API rate. Below roughly 150 images a month, pay-as-you-go usually wins. Above that, a subscription almost always does.
Watch one trap specific to this migration. Because Nano Banana returns a single image per call, a workflow that used to request four variations in one Imagen call now makes four billable calls. Your per-image rate did not change, but your bill can quadruple without anyone editing a prompt.
How do you test this in the next twenty minutes?
Run one prompt through three models and score the output on the axis you actually care about. Twenty minutes of side-by-side testing on your own brief tells you more than any comparison article, including this one, because your subject matter and quality bar are not the ones the reviewers used.
The exercise:
--- Take a real brief you shipped in the last month, not a test prompt
--- Run it through Nano Banana, Midjourney and one of FLUX.2 or GPT Image 2 using the template above, unchanged across all three
--- Score each output out of five on three things only: did it follow the brief, is the lighting believable, and how many rounds of editing would it still need
--- Then take the best output and try to edit one element out of it in each tool
That last step is the one people skip, and it is where the ranking usually flips. A model that generates a slightly weaker first image but edits cleanly will beat a prettier model that has to be re-rolled from scratch every time the client asks for one change.
Keep the scores in a note. When the next model ships, and something ships roughly every six weeks now, you will have a baseline instead of a vibe.
What should you do this week?
Audit where Imagen is hiding in your stack, rewrite your negative prompts as positive descriptions, and re-test your model routing against a real brief. Those three moves take an afternoon and they prevent the two failure modes of this migration: silent breakage and silently worse output.
The wider lesson is that model deprecation is now a normal operating cost of working with AI, not an unusual event. Imagen went from flagship to shut down in under two years, and Gemini 2.5 follows in October. Building on a named model without a note of what replaces it is the modern version of not backing up your files.
None of that is a reason to slow down. It is a reason to keep a one-page record of which model each of your workflows calls, so a deprecation notice becomes a fifteen-minute task instead of a week of confused debugging.
Reliability, not novelty, is what separates people who get value from AI from people who keep starting over. We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth.
If you want more on making AI output behave consistently, our guide to context engineering for non-coders covers the text-side equivalent of everything above, and our breakdown of AI video plan limits applies the same cost maths to video.
Reviewed by the UD AI team.
Put the Right Model on the Right Job
Knowing which model wins each task is one thing. Building it into a workflow your whole team runs the same way every time is another. We'll walk you through every step, from model selection and prompt templates to the automation that connects them.
