Most people whose work depends on Google's image generation do not know that three of the endpoints they rely on stop existing in six days. There is no warning banner in the tool you are using. There is no email. On 17 August 2026, Google shuts down every Imagen 4 endpoint, and anything still pointing at them fails.
If you generate images through Google inside an automation, a template, a Sheets script, or a third-party app that resells Google's image models, this is your problem and not your developer's. The fix takes about twenty minutes. Waiting does not make it smaller.
What exactly is Google shutting down on 17 August 2026?
Google is retiring the three Imagen 4 endpoints in the Gemini API: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001. All three were released on 24 June 2025. The recommended replacement for all three is gemini-3.1-flash-image, the model most people know by its nickname, Nano Banana.
A shutdown is not a deprecation. Google's own wording separates the two: a deprecation is the announcement that support has ended, and a shutdown is the point at which the endpoint is completely turned off and no longer available.
The dates come from Google's Gemini deprecations page, last updated 3 August 2026. Google adds one caveat worth reading carefully: the listed dates are the earliest possible shutdown dates, and it says it will confirm the exact date with advance notice. Plan for 17 August anyway. A migration you finished early costs you nothing.
The facts you can act on
--- Shutdown date: 17 August 2026 (earliest possible, per Google)
--- Affected: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001
--- Replacement: gemini-3.1-flash-image (Nano Banana)
--- Also shutting down this month: embedding-2-preview (10 August 2026) and gemini-robotics-er-1.6-preview (31 August 2026)
--- Not affected: gemini-3.1-flash-image and gemini-3-pro-image, both released 28 May 2026, with no shutdown date announced
How do I check whether my image workflow is affected?
Search every place a Google model name could be stored for the string imagen-4. If it appears anywhere, that path breaks. Most practitioners have between one and four such places and forget at least one of them, usually a Zap or scenario built months ago and never opened since.
Here is the full list of places to look, in the order that catches the most problems fastest.
--- Your Make.com scenarios and n8n workflows: open every Google AI or Gemini module and read the model field, not the module name
--- Your Zapier Zaps: check the Google AI Studio and Gemini actions
--- Google Apps Script and Sheets add-ons: search the script for imagen
--- Any custom GPT, agent, or internal tool a colleague built for you
--- Third-party apps that let you paste your own Google API key: the model is often buried in an advanced settings panel
If you cannot read your own automation config confidently, paste it into an assistant with this prompt. It works because it forces a yes or no answer on a specific string rather than a general opinion about your setup.
Try this prompt
You are auditing an automation config for a hard API deprecation. Below is the raw configuration of my workflow. Google shuts down these three model IDs on 17 August 2026: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001.
Do three things and nothing else:
1. List every location in the config where any of those three strings appears, quoting the surrounding field name.
2. For each one, state the exact field I need to change and the value to change it to (gemini-3.1-flash-image).
3. List any other Google model IDs you find, and mark each one "check" so I can look it up myself. Do not guess whether they are affected.
If none of the three strings appear, reply only with: NOT AFFECTED. Do not reassure me, do not summarise my workflow, and do not suggest improvements.
CONFIG:
[paste your config, scenario JSON, or script here]
What actually changes when you move from Imagen 4 to Nano Banana?
Three things change, and only one of them is the model name. The method changes, the request and response shape change, and the way the model reads your prompt changes. Swapping the model ID alone leaves you with a broken call or, worse, images that look nothing like the ones you were producing last week.
The method is different. Imagen 4 was called through a dedicated image method. Nano Banana is not a separate image endpoint at all. It runs through the same general content generation call as text, which is why the request and response you send and parse both change shape.
Settings move. Aspect ratio, number of images, and similar controls no longer sit where they used to. On Gemini image models these live inside the generation configuration, alongside a declaration that you want an image back rather than only text. If your automation platform exposes these as tick boxes, check them after you switch. Defaults are rarely the ones you had.
Your prompts behave differently. This is the part that catches people who did the technical swap correctly and still ended up with worse output. Imagen 4 rewarded dense, comma-separated descriptor stacks. Gemini image models are conversational: they respond better to a described scene in plain sentences, and they hold a conversation about an image across turns, so you can ask for an edit instead of regenerating from scratch.
Google publishes the mechanics of the swap in its Imagen to Gemini image migration guide. Read the prompting notes, not just the code.
How do I rewrite my Imagen prompts for Gemini 3.1 Flash Image?
Stop writing keyword stacks and start writing sentences. An Imagen 4 prompt that read like a tag list will still produce an image on Nano Banana, but a flatter and less controlled one. Rewriting takes one pass per prompt template and is the single highest-return part of this migration.
The pattern that works: name the subject, then describe the scene as you would to a photographer, then state the constraints last. Constraints at the end survive better than constraints buried in the middle.
Before, written for Imagen 4
professional headshot, asian businesswoman, 30s, navy blazer, office background, bokeh, natural light, high detail, 4k, sharp focus
After, written for Gemini 3.1 Flash Image
A professional headshot of a Hong Kong businesswoman in her thirties, wearing a navy blazer. She is standing in a modern open-plan office; the background is softly out of focus. Warm natural daylight from a window on her left. Shot at eye level, waist up, with the subject positioned slightly right of centre.
Keep it photorealistic. No text, no watermark, no logo. 16:9.
The second version is longer and that is the point. You are giving the model a scene to construct rather than a bag of adjectives to average.
One habit worth building while you are in there: because Gemini image models keep context across turns, your second prompt should be an edit, not a rewrite. "Same image, but change the blazer to charcoal grey and keep everything else identical" is faster and more consistent than regenerating. This is the same consistency problem practitioners hit in other tools, and we covered the prompt-side fix for it in why your Midjourney images never match.
What breaks if you wait until 17 August?
The failure is loud and immediate, not gradual. A retired endpoint returns an error rather than quietly routing you to a newer model, so an automation that runs unattended does not produce worse images on 18 August. It produces no images, and depending on how the scenario is wired, it may fail silently from your point of view because nobody is watching the run log.
Three specific ways this bites practitioners.
--- Scheduled content pipelines. If your Monday social batch generates images automatically, you find out when the post goes out with a blank slot, or does not go out at all
--- Client-facing tools. Anything you built for a client that calls Imagen 4 becomes their outage and your phone call
--- Templates you shared. Scenario templates and shared workflows carry the old model ID to everyone who copied them, and none of those people are reading Google's deprecation page
There is also a version of this that is worth being honest about: migrating early has a real cost. Your prompt library needs rewriting, your image style will shift slightly even after a careful rewrite, and if your brand approvals depend on visual consistency, you need a review round before the switch rather than after. That is an argument for starting today, not for waiting.
Try it now: the twenty-minute migration check
Set a timer for twenty minutes and work through five steps in order. Most practitioners finish inside the window; the ones who do not are the ones who discover a workflow they had forgotten about, which is exactly the point of doing this before the deadline rather than after.
--- Minutes 0 to 5. Open every automation that touches image generation and search the config for "imagen-4". Write down each hit
--- Minutes 5 to 8. Change each model field to gemini-3.1-flash-image. Save
--- Minutes 8 to 12. Check the generation settings on each one: aspect ratio, output count, and whether the module is set to return an image and not only text
--- Minutes 12 to 17. Take your single most-used prompt, rewrite it from keyword stack to described scene using the pattern above, and run it once
--- Minutes 17 to 20. Compare the new output against last month's version side by side. If the style has drifted, adjust the constraint line at the end of the prompt rather than the whole prompt
If you get a "NOT AFFECTED" in step one, you have lost nothing and gained an inventory of where your image generation actually lives. That inventory is worth having the next time a model gets retired, which on the current cadence will be within a few months.
The real lesson: model retirement is now part of your workflow
Model shutdowns are a permanent feature of working with AI, not a one-off inconvenience. Google's deprecations page currently lists shutdown dates for more than a dozen models across image, video, embedding, audio and robotics. Anthropic and OpenAI publish equivalent notices. Nobody sends you a personal email when yours is on the list.
The practitioners who stay ahead of this do one small thing: they keep a written list of every model ID their workflows depend on, and they check the vendor deprecation pages once a month. It takes ten minutes and it converts a fire drill into a calendar entry.
The other half of the lesson is that a migration is never only technical. The endpoint swap takes three minutes. Rebuilding the prompt library so the output still looks like your brand is the actual work, and it is the part that gets skipped under deadline pressure. Give it the review round it deserves.
Technology moving this fast is genuinely tiring, and pretending otherwise does nobody any favours. We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth. When a vendor turns something off with six days' notice, having someone who has already read the changelog matters more than having the newest tool.
Rebuild It Once, Properly
You have the migration steps. The harder question is which model your workflow should be standing on next time, and how to build it so the next shutdown is a five-minute change instead of a scramble. We'll walk you through every step, from choosing the model to rebuilding the prompt library and wiring the automation so it fails loudly instead of silently.
Reviewed by the UD AI team. Model IDs and shutdown dates verified against Google's official Gemini deprecations page on 11 August 2026. Shutdown dates published by Google are earliest-possible dates and may move; check the source page before relying on them.
