Your image workflow runs on a model Google scheduled for shutdown in June

The model nobody in your business chose

Somewhere in the workflow that turns a product description into a social post, there is a field called Model. If you built that workflow in n8n after April, you almost certainly never touched it. It arrived filled in.

The value it arrived with is models/gemini-3.1-flash-image-preview. Google announced a shutdown date for that model of 25 June 2026. That date passed more than three months ago.

Your workflow may well still be running. That is the confusing part, and it is the reason this is worth ten minutes today rather than an emergency later. What follows is what the vendor has actually committed to, what the failure looks like when it comes, and why updating your platform does not change the value sitting in the field.

What "shutdown" means here, and how firm the date is

Google keeps a deprecations page for the Gemini API, and it defines its own terms:

A "deprecation" is the announcement that we no longer provide support for a model, and that it will be "shut down" in the near future. Once a model is "shutdown", it is completely turned off, and the endpoint is no longer available.

• Gemini API deprecations, Google, page updated 5 September 2026

The announcement for this particular model is in the changelog, dated 28 May 2026:

Deprecation announcement: The gemini-3.1-flash-image-preview and gemini-3-pro-image-preview models are deprecated and will be shut down on June 25, 2026.

• Gemini API changelog, Google

Twenty-eight days of notice, on a developer changelog, for a model that the average person running the workflow has never heard of and did not select.

The date itself is softer than it looks, and Google says so in a note on the deprecations page:

The shutdown dates listed in the table indicate the earliest possible dates on which a model might be retired. We will communicate the exact shutdown date to users with advance notice to ensure a smooth transition to a replacement model.

• Gemini API deprecations, same page

So the honest description of your workflow's state is not "it broke in June". It is that the model has been past its earliest possible retirement date since June, that the only date the vendor has published has already gone by, and that the further notice it promises has no minimum length and no address on it. The last one arrived on a developer changelog.

Is it actually switched off right now?

Probably not, and the evidence matters here, because a claim that everything is already broken is easy to disprove and would be wrong.

Google marks retired models on the deprecations page: "Already-shutdown models are indicated with gray backgrounds". Thirty-two rows on that page carry the grey marker. The row for gemini-3.1-flash-image-preview does not, even though its date has passed. Nor do the rows for the three Imagen 4 models with an announced shutdown of 17 August 2026.

There is a second signal, and it is stronger. When Google actually pulls a model, it posts a confirmation in the changelog, in a consistent form: "The gemini-3.1-flash-lite-preview model has been shut down", "The following Gemini 2.0 models are now shut down", "The gemini-2.5-flash-image-preview model has been shut down". For the 25 June and 17 August image models there is no such entry, in any changelog post since.

And there are two observations from Google's own developer forum. On 15 July, twenty days after the announced date, somebody described a production platform running on gemini-3.1-flash-image-preview and complained about output quality rather than about a 404. On 3 September, another developer reported that listing models through the API still returns every image model including both previews.

What it looks like when it does stop

This exact failure has happened before, to the model that was retired in January, and the error text is the part to remember. Here is what a user saw in n8n:

"errorMessage": "The resource you are requesting could not be found"

"errorDescription": "models/gemini-2.5-flash-image-preview is not found for API version v1beta, or is not supported for generateContent. Call ListModels to see the list of available models and their supported methods."

• In n8n, the Edit Image node throws a 404, n8n community, 16 January 2026

The first line is what most people see, and it says nothing about models. "The resource you are requesting could not be found" reads like a missing file, a wrong account, an expired credential. Everything you would check first is the wrong thing to check. The actual cause is in the second line, which is a level deeper than most error panels show.

The same wording has shown up in three separate reports, on two platforms and with two different models — the n8n forum in January 2026, a GitHub issue five days later about the same model, and the Make community back in May 2025 with gemini-2.0-flash-preview-image-generation. It is Google's standard response for a model that is no longer served, and it will be the same sentence with a different model name when this one goes.

One more thing to know for the diagnosis: n8n's node has an error of its own that looks superficially similar, Model … is not supported for image generation. That one comes from the node, not from Google, and it only fires when the model name contains neither "gemini" nor "imagen". If your model name looks right and you are getting a 404 anyway, the answer is on Google's side.

Why updating n8n will not move the default

The default was set on 23 April 2026, in a pull request whose description lists the choice in one line: "Image Generate -> models/gemini-3.1-flash-image-preview". Google announced the model's deprecation thirty-five days later, and its shutdown date fell sixty-three days after the default was set.

Since then the file has been touched twice — on 27 April, to show an option only for the models that support it, and on 20 May, for an unrelated fix to prompt validation. Other files in the same node have been updated repeatedly through the summer. The default model has not changed, and there is no open issue in the n8n repository asking for it to change.

The mechanism is worth understanding, because it explains how "I never touched it" and "it is running a retired model" can both be true. The model dropdown is live: the node fetches the list of models from the API each time you open it. The default is not. It is a fixed string, applied once, when you add the node, and then saved inside your workflow. If the model later disappears from the live list, your saved value does not disappear with it. It just keeps being sent.

This is the same shape as a model name pinned in a dropdown that still looks current, and the same thing happened to a voice workflow whose model was chosen for it. The common factor is that the choice was made by the tool, at a moment nobody remembers, and the tool has no obligation to revisit it.

There is an irony in the record here. In January, when the previous hardcoded image model was retired, somebody filed a bug about it. It was closed as working as expected, with this reply:

Note: As you are still running on an old version (1.122.4 while we are on 2.4.4 currently), it is expected that as AI providers update their models and APIs, things will break. Please update to the latest version and you will not only find that there is a model selector, the error also won't exist.

• n8n issue #24658, 22 January 2026

Three months after that advice was given, the update it recommended shipped a new hardcoded default pointing at a model Google would announce for shutdown in another five weeks. Neither statement was wrong when it was made. That is exactly the problem: "keep your platform updated" is good advice that does not solve this, and believing it does is how a workflow ends up on a retired endpoint.

Google's replacement table points at a model already past its shutdown date

If you go to the deprecations table to find out what to use instead, it will send you somewhere strange. Here are two rows from that page, out of two different sections of it, side by side:

The replacement offered for the model dying tomorrow is a model whose own shutdown date, printed higher up the same page, was ninety-nine days earlier.

Meanwhile, Google's image generation guide — a different page, updated the day before the deprecations page — gives entirely different advice for the same model:

Nano Banana (Gemini 2.5 Flash Image) (gemini-2.5-flash-image): The legacy pioneer of the Nano Banana series. While it has been a reliable workhorse, we strongly recommend that customers transition to Nano Banana 2 Lite to experience enhanced quality, faster generation speeds, and lower API pricing.

• Image generation, Google, page updated 4 September 2026

Two vendor pages, updated within a day of each other, recommending two different successors — and the one the table names is the model this article is about, past its own announced shutdown date since June. Follow the table and you migrate from one expired endpoint onto another. The generally available models — the ones with no announced shutdown date at all — are gemini-3.1-flash-image, gemini-3.1-flash-lite-image and gemini-3-pro-image. Those are the names worth typing.

Tomorrow's date, for everyone who already moved once

There is a live deadline attached to this, and it is 2 October 2026: the announced shutdown of gemini-2.5-flash-image, the model many people switched to when the January retirement broke their workflows.

That one is a generally available model rather than a preview, released on 2 October 2025, exactly a year before the date now set for its shutdown. It is also the value most likely to be sitting in a Make scenario or a Zapier step right now, because it was the current model when those were built. The same caveat applies as everywhere else on that page — the date is the earliest possible one — but it is the nearest date on the calendar, and unlike the June one it has not passed yet.

Make and Zapier freeze the choice in a different place

Neither platform has n8n's hardcoded default, and both end up in a similar position by a different route.

On Make, image generation through Gemini does not run through the module you would expect. A Make employee explained it in the community:

Nano banana is an available model within the [Gemini] app. One wrinkle — as Nano Banana is designed to be prompted from a chat, it's treated in the app as a text model. You'll need to use the "Generate a response" module rather than "Generate an image".

• a Make employee, How can I integrate Nano Banana Pro with Make?, Make community, 13 February 2026

In the same thread, the instruction for picking a model was "First, select a Nano Banana model (Gemini 2.5 Flash Image or Gemini 3 Pro Image Preview)". One of those two names is the model dying tomorrow, and the other is a preview whose shutdown was announced for June. A scenario built from that advice in February has one of those two ids frozen in it — either the model dying tomorrow or the preview announced for shutdown in June — and Make's app documentation does not publish which models the Gemini modules support, so there is nothing to check against.

On Zapier, the Generate Image action takes the model as a required field with one sentence of help: "The model to use for image generation. Supports Gemini and Imagen models." Whatever was set there once is what runs until somebody opens the Zap again. The neighbouring video action goes further and suggests preview model ids directly in its help text.

Across all three platforms, no vendor has published a warning about these retirements to the people using the integrations. Zapier's help centre has nothing about Gemini model deprecation at all. Make has published no release note about the Gemini image models. n8n's own documentation describes the Generate an Image operation without naming the model it defaults to. The notice exists in a developer changelog, and nowhere a business owner would pass by.

What the swap costs

Changing model changes your bill, and the direction depends on which replacement you pick. Google's pricing page, per 1024-pixel image:

Moving to the flagship — which is what "pick the newest one" gets you — costs about seventy per cent more per image. Moving to the Lite model, which is what Google's own guide recommends, costs about a seventh less than what you pay today. On a workflow generating a few hundred images a month the difference is small in absolute terms; on one generating thousands, it is the kind of change that shows up on a card statement without any explanation attached to it.

None of the three has a free tier. If your workflow has been running on free quota, it is not running on any of these.

What to do this week

1. Open every workflow that generates or edits images and read the Model field out loud. You are looking for anything with preview in the name, and for gemini-2.5-flash-image.
2. Check the Edit Image operation separately if you use it. It has no default at all, so whatever is in there was chosen by a person, on a day when the list looked different.
3. Set the model explicitly to a generally available name — models/gemini-3.1-flash-image, models/gemini-3.1-flash-lite-image or models/gemini-3-pro-image. The field accepts a typed id as well as a picked one, so you do not have to wait for a dropdown to refresh.
4. Run it once and look at the picture, not the status. A different model is a different look. If the output feeds a client-facing template, somebody should see the new output before a customer does.
5. Do the same pass in Make and Zapier, where the value is saved per module and per Zap. There is no central place that lists them.
6. Write the model name into whatever documentation you keep. The reason this became an incident twice is that nobody could say what the workflow was using without opening it.
7. Do not rely on the vendor's replacement column. Check the model list page for what has no shutdown date, which is the only field on any of these pages that answers the question you actually have.

When this is a job to hand over

Three workflows, one field each, fifteen minutes with a coffee. That is genuinely the whole job for most people reading this, and the reason to do it now is that it costs the same today and considerably more on the morning it fails.

The version that is not that has a specific shape. Images that go out under a client's brand need the new model reviewed rather than just switched, because the output changes and nobody signed off on the new look. Agencies with the same template deployed across a dozen accounts have a dozen copies of the same stale string and no inventory of where they are. And any workflow where the image feeds a scheduled post rather than a person means the failure surfaces as a missing post, noticed a day late, by someone who does not have access to the automation.

Those are countable — how many workflows, how many accounts, how many templates — and countable is what makes a fix quotable before anyone starts. The AI workflows we take on are frequently exactly this: something the platform filled in by default, two releases ago, that quietly stopped being a good idea.

Most AI workflows contain at least one value nobody chose deliberately, sitting in a field nobody opens, pointing at an endpoint with a published expiry date. The image model is simply the one with a date on it this week. Finding the others is the same exercise, and it is best done while somebody is actually watching the runs rather than after the first blank post goes out.

Sources

• Gemini API deprecations - Google, page updated 5 September 2026, read 7 September 2026. The definitions of deprecation and shutdown, the note that listed dates are the earliest possible ones, the grey-background convention for already-retired models, the announced dates for the image preview models and Imagen 4, and the replacement column pointing gemini-2.5-flash-image at a model retired in June.
• Gemini API changelog - Google, read 7 September 2026. The 28 May 2026 deprecation announcement for both image preview models, the 15 June announcement for the Imagen 4 models, and the pattern of separate confirmation entries when a model is actually shut down, which has not appeared for either of these.
• Image generation - Google, page updated 4 September 2026, read 7 September 2026. The recommendation to move from Gemini 2.5 Flash Image to Nano Banana 2 Lite.
• Gemini API models - Google, page updated 4 September 2026, read 7 September 2026. The list of current image models, from which both preview models are absent, and Imagen 4 marked as deprecated.
• Gemini API pricing - Google, page updated 4 September 2026, read 7 September 2026. The per-image prices for the three current image models and the absence of a free tier for any of them.
• In n8n, the Edit Image node throws a 404 - n8n community, 16 to 17 January 2026, read 7 September 2026. The full error payload including the two-line message, and a second user in the same thread reporting the same problem.
• n8n issue #24658 - n8n on GitHub, 21 to 22 January 2026, read 7 September 2026 through the GitHub API. The January break caused by a hardcoded model id, the closure as working as expected, and the reply advising an update.
• n8n pull request #28853 - n8n on GitHub, merged 23 April 2026, read 7 September 2026 through the GitHub API. The pull request that set the current default image model, with its description naming the model explicitly.
• n8n source, Google Gemini node - read 7 September 2026 from raw source and the commits API. The hardcoded default on the Generate an Image operation for node versions 1.2 and above, the absence of any default on Edit Image, the live model list fetched from the API, the node's own unsupported-model error and when it fires, and the commit history showing the default untouched since it was set.
• How can I integrate Nano Banana Pro with Make? - Make community, 13 and 16 February 2026, read 7 September 2026. A Make employee explaining that Gemini image generation runs through the Generate a response module, and the instruction to select either Gemini 2.5 Flash Image or Gemini 3 Pro Image Preview.
• Google AI Studio (Gemini) on Zapier - Zapier Help Center, updated 10 August 2026, read 7 September 2026, together with the published field definitions for the Generate Image action. The model as a required free-text field with no default and no live list, and the absence of any Zapier article about Gemini model deprecations.
• Gemini image quality after iterative edits and image model availability on the OpenAI-compatible endpoint - Google AI developer forum, 15 July and 3 September 2026, read 7 September 2026. A production platform still running on gemini-3.1-flash-image-preview three weeks after its announced shutdown date, and a September report in which the API still lists both preview models.