Google Shut Down Three Imagen 4 Endpoints Yesterday. If Your Side Project Still Calls imagen-4.0-generate-001, It Is Already Broken.
Google's own model page carries the notice in plain text: the Imagen 4 standard, ultra, and fast endpoints are deprecated and were shut down on August 17, 2026. That was yesterday. If you have anything still calling imagen-4.0-generate-001, imagen-4.0-fast-generate-001, or imagen-4.0-ultra-generate-001, it is not degraded or slower. It is off.
I found one of mine at 9am. A thumbnail generator for an old project, four lines, hardcoded model string, last opened in February. It had been failing silently into a try/except for however many hours, and the only reason I caught it is that I went looking specifically because of this news.
The migration is not a string swap
The part that makes this more annoying than a normal deprecation is that the replacement is not Imagen 5. Google's notice points you at Gemini 3.1 Flash Image, which is a general Gemini model that happens to generate images, not a successor in the Imagen line.
That distinction matters because Imagen had its own call shape. It was a dedicated image endpoint: you passed a prompt, you got back one to four images, and the request looked nothing like a text completion. Moving to a Gemini image model means image generation now lives in the same call path as everything else you send to Gemini. Reports from people who have done the migration describe the dedicated image-generation method going away entirely in favour of the standard content-generation call.
I have not migrated a production Imagen integration myself, so treat my characterisation of the new call shape as secondhand. What I can say with confidence from the official page is that the target is a different model family, and different model families do not have identical interfaces.
On price, I have seen a figure of roughly $0.067 per image quoted for the replacement against a range of about $0.02 to $0.06 for the Imagen 4 tiers. I could not confirm those numbers against Google's own pricing page, so do not budget off them. Check the current pricing yourself. The directional claim, that the replacement is not cheaper, seems well supported. The specific cents are not.
The Vertex side already happened, which is why some people got hit twice
There were two separate tracks here. The Gemini API deprecation is the August 17 one. The equivalent Imagen 4 endpoints on the Vertex AI side reportedly hit their deprecation date earlier in the summer.
If you run one project on the Gemini API and another on Vertex, you got two separate breakages on two separate dates for what you reasonably thought of as one model. That is not Google being sloppy. It is what happens when the same model is shipped through two products with two release calendars, and it is a good argument for knowing which surface each of your projects is actually talking to.
The real cost is not the API bill
Here is the thing I keep relearning. The expensive part of a deprecation is almost never the price delta. It is that you have to open a file you have not thought about in eight months, remember how it works, verify the new output is acceptable, and redeploy something that was not on your list this week.
For a company with an engineering team, that is a ticket. For one person, that is your Tuesday.
And the Imagen 4 page tells you exactly why this class of thing keeps happening: the model's listed last update is June 2025. Fourteen months from last update to shutdown is a completely normal lifecycle for a hosted model right now. If you are building on hosted AI endpoints, you should plan on roughly annual forced migrations per model you depend on, and price that into whether a given integration is worth having at all.
The ten-minute audit
This is what I did after finding the broken thumbnail generator, and it is the only part of this post I would actually ask you to act on.
Grep every repo you own for hardcoded model identifiers. Not just Google ones. In practice this is one command per provider prefix:
rg -n --no-heading \ -e 'imagen-[0-9]' \ -e 'gemini-[0-9]' \ -e 'gpt-[0-9]' \ -e 'claude-[a-z0-9-]*-[0-9]' \ -e 'text-embedding-' \ ~/Projects
Then do three things with the output.
Move every match into config. A model ID scattered across four files is four edits under time pressure. In one variable it is one edit.
Write down, somewhere you will actually look, which projects depend on which provider surface. Gemini API and Vertex AI are different answers even for the same model name.
Make failures loud. My thumbnail generator swallowed the error because past me wrapped it in a bare except and moved on. A model endpoint returning 404 should page you or at least show up in a log you read, because the whole failure mode here is silence.
If you want the version that takes an extra twenty minutes: subscribe to the deprecation pages rather than the release blogs. Google publishes a dedicated deprecations page for the Gemini API. Providers announce shutdowns there long before the date, and almost nobody reads them, which is precisely why the date arrives as a surprise.
Where I could be wrong
I am framing this as a maintenance-tax story, and there is a reasonable counter-position: hosted model churn is the price of not running your own inference, and it is a very low price. Fourteen months of a managed, scaled, patched image endpoint for the cost of one afternoon of migration is a good trade compared to operating that yourself. I mostly agree with that.
The place the trade goes bad is when the migration is not one afternoon, and the way you end up there is by having many shallow integrations across many providers instead of a few deep ones. Every additional provider you touch adds its own deprecation calendar to your life. That argues for consolidating on fewer surfaces rather than for self-hosting.
The other thing I could be wrong about is urgency. If you checked your projects and nothing calls Imagen 4, this post costs you nothing and you can close it. That is the likely outcome for most people reading, and I would rather you run the grep and find nothing than skip it and find out in November.
Author
Lukas
@lukcombinator