Yesterday, Google’s Gemini API deprecations page quietly did what it said it would do back on June 15: it shut down imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, and imagen-4.0-fast-generate-001. Not throttled. Not deprecated-with-a-warning-banner. Shut down. Any app still calling those model IDs on August 17, 2026, started returning errors instead of images, and the method those apps were built on — generate_images() — doesn’t exist anymore either. Google didn’t retire a model here. It deleted a code path.

That distinction matters more than it sounds. A retired model usually means swapping a string. This one means rewriting how your app talks to the API at all, because Google folded image generation into the same generate_content() interface it uses for text, and the replacement model has a different response shape, different parameters, and — per early developer reports — different output quality.

Quick Summary: What Happened

Detail	Info
Shutdown date	August 17, 2026
Models killed	imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001
What broke	The generate_images() method was removed entirely — a hard error, not a warning
Replacement	Gemini 3.1 Flash Image, marketed as “Nano Banana 2,” called via generate_content()
Notice given	Deprecation announced June 15, 2026 — roughly nine weeks before shutdown, per Kingy AI’s launch tracker
Official source	Google Gemini API deprecations

Bottom line: If your app called generate_images() and you hadn’t migrated by yesterday, it’s throwing errors in production right now — and the fix isn’t a one-line swap.

---

What Actually Happened

Google’s own deprecations documentation lists all three Imagen 4 GA endpoints — standard, ultra, and fast — with the same shutdown date: August 17, 2026. That’s not new information; it’s been sitting on that page since mid-June. What’s new is that the date arrived, and per byteiota’s migration coverage, Google didn’t ease anyone into it. The generate_images() method — the call thousands of apps used to request Imagen output — is gone from the SDK. Call it today and you get a hard error, not a deprecation notice logged to your console.

That’s a genuinely unusual choice. Most API sunsets leave the old method in place returning a warning for a stretch before pulling it. Google skipped that step here. Which means the population of affected developers isn’t “everyone who uses Imagen” — it’s specifically everyone who didn’t read the deprecations page in the nine weeks between June 15 and yesterday. Judging by the Google Developer forums, that’s a meaningful number of people.

Vertex AI users had a preview of this. Google had already been winding down GA Imagen access on that platform ahead of the Gemini API’s own timeline — the same pattern of “the model still technically works until, abruptly, it doesn’t.” Anyone running Imagen on Vertex AI who watched that happen and still hadn’t touched their Gemini API integration doesn’t have much of an excuse left.

What Actually Breaks When You Call generate_images() Now?

1. The method call itself. generate_images() no longer exists in the SDK — calling it raises an error instead of returning a deprecation warning.
2. The response object. Gemini 3.1 Flash Image returns images nested inside response.candidates[0].content.parts, not as a clean array of image objects the way Imagen did.
3. Batch generation. The number_of_images parameter is gone. Generating multiple variants now means looping individual generate_content() calls instead of one batched request.
4. Prompt controls. Negative prompts and some format controls that Imagen supported aren’t carried over to the new interface.
5. Watermarking. SynthID watermarking is mandatory on Gemini 3.1 Flash Image output and can’t be disabled — a change worth knowing about if your use case required clean output.

None of these are cosmetic. A team that built error handling, retry logic, or downstream image processing around Imagen’s response shape has actual engineering work ahead, not a config change.

Why This Matters

Two separate problems are colliding for developers this week, and it’s worth pulling them apart because they call for different responses.

The first is the outage itself. If generate_images() was load-bearing in your production app, you have a live incident, not a maintenance item. That’s a today problem.

The second is what you migrate into, and this is where it gets more complicated than “swap the model name.” Gemini 3.1 Flash Image — Nano Banana 2 — is priced by output resolution rather than by a flat per-image rate: $0.045 for a 512px image, $0.067 at 1024×1024, up to $0.151 at 4K. Imagen 4 Standard, by comparison, was a flat $0.04 per image regardless of size. At the resolution most apps actually use — 1K — that’s roughly a 67% cost increase per image over what Imagen 4 Standard charged. Google’s batch API cuts output pricing in half, which helps for non-interactive workloads, but if your app generates images synchronously in response to a user action, batch pricing usually isn’t available to you.

Then there’s quality. On the Google Developer forums, one developer building a portrait-generation workflow described the difference bluntly: Imagen 4 delivered “high-fidelity, highly consistent results,” while the Nano Banana replacement produces “heavy semantic drift — the facial features vary wildly across the 4 grids, often producing entirely different people.” The same thread flags a loss of skin texture and a shift toward an “overly plastic/illustration-like” look compared to Imagen 4’s photorealism. That’s not one disgruntled poster — similar complaints about over-smoothed, plastic-looking output show up across multiple threads on Google’s own forums this year. If your app depends on consistent character or face rendering across multiple generations, this is the thing to test before you assume the migration is a drop-in win.

What Are Your Options Now

If you’re still on generate_images(), that’s the fire to put out first. There’s no version pin or grace-period flag to buy time — the method is gone, not slow. Rewrite the call using generate_content() against gemini-3.1-flash-image and update your response parsing to pull images out of response.candidates[0].content.parts.

Budget for the pricing shift before you ship the migration, not after. Model your actual output resolution against the new per-image rates rather than assuming cost parity with Imagen 4. A 1K-heavy workload just got materially more expensive per call.

Test consistency-sensitive use cases specifically. If your product does anything with recurring characters, portraits, or brand-consistent visual assets, run a side-by-side batch before you cut over. The regressions developers are reporting cluster around exactly that kind of output.

If Nano Banana 2’s output doesn’t clear your bar, this is a reasonable moment to shop around rather than force a Google-to-Google migration. Our guide to the current AI image generator field covers alternatives worth testing against your specific use case before you commit engineering time to one vendor’s replacement model.

Read the deprecations page as a standing habit, not a one-time check. Google didn’t hide this — it sat on a public page for nine weeks. The lesson isn’t “Google was sneaky.” It’s that a URL nobody has bookmarked isn’t a reliable alert system for a hard production dependency.

The Bigger Picture

This is the second time in two weeks we’ve written about a piece of AI infrastructure quietly turning into a reliability risk that has nothing to do with model intelligence. We covered how a compromised build pipeline let backdoored packages into LiteLLM and exposed thousands of companies through an unaudited gateway dependency, and days after that, Stripe paid a 5x premium to buy OpenRouter — a bet that owning the routing layer between apps and models is worth real money precisely because that layer is where operational risk actually concentrates. Imagen 4’s shutdown is a smaller story than either of those, but it’s the same category: the part of the AI stack that breaks isn’t usually the model’s reasoning. It’s the connective tissue — the method signature, the response shape, the deprecation calendar — that most teams don’t monitor with the same discipline they apply to the model itself.

It’s also worth naming that Google is shipping these changes fast. Gemini 3.1 Pro landed in February with a genuinely large reasoning jump, and its image sibling has already gone through a public naming and quality debate in under six months. Rapid iteration is a real strength when the model is improving. It’s a real cost when the API contract underneath your app changes shape on the same schedule.

Our Take

We don’t think Google did anything indefensible by shutting Imagen 4 down. Nine weeks of public notice on an official deprecations page is a normal runway for an API sunset, and folding image generation into generate_content() is a defensible architectural decision — one interface instead of two is genuinely simpler to maintain, for Google and eventually for developers too.

What we’d push back on is treating this as fully resolved once the migration code compiles. The response-shape rewrite is the easy part. The harder, less visible cost is that Nano Banana 2 isn’t a straight quality upgrade for every use case Imagen 4 covered — the portrait-consistency complaints on Google’s own forums are specific and repeated enough to take seriously, and the resolution-based pricing means “migrated” and “cost-neutral” aren’t the same claim. If you’re the engineer who just got the production alert, finish the fix. But before you mark this closed, actually compare output quality against your old Imagen 4 results side by side. The forum threads suggest a decent number of teams that skipped that step are going to be back in this migration a second time.

Frequently Asked Questions

Is Imagen 4 completely shut down now?

Yes. As of August 17, 2026, the three GA Imagen 4 endpoints — imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, and imagen-4.0-fast-generate-001 — are shut down on the Gemini API, per Google’s own deprecations page. The generate_images() method used to call them was removed entirely, not just deprecated.

What replaces Imagen 4?

Google’s recommended replacement is Gemini 3.1 Flash Image, marketed as “Nano Banana 2.” Unlike Imagen, it’s called through the generate_content() method — the same interface used for text generation — rather than a dedicated image-generation method.

Why did my app break instead of showing a warning first?

Google removed the generate_images() method outright rather than deprecating it with a grace period. Per byteiota’s coverage, calls to the old method now raise a hard error. There was no interim warning state — the deprecation notice lived on Google’s documentation page, not in the API response itself.

Is Gemini 3.1 Flash Image more expensive than Imagen 4?

For most common resolutions, yes. Imagen 4 Standard was a flat $0.04 per image. Gemini 3.1 Flash Image charges by output resolution: $0.045 at 512px, $0.067 at 1024×1024, and up to $0.151 at 4K. At the 1K resolution most apps use by default, that’s roughly a 67% increase per image. Google’s batch API cuts output cost in half, but it isn’t available for synchronous, user-triggered generation.

Does Nano Banana 2 produce lower-quality images than Imagen 4?

Results vary by use case, but developers on Google’s own forums have specifically reported quality regressions in portrait generation — describing “semantic drift” where facial features shift unpredictably across generations, along with a loss of skin texture compared to Imagen 4’s output. Test your specific use case before assuming parity.

What code changes do I need to make to migrate?

Replace generate_images() calls with generate_content() targeting gemini-3.1-flash-image, and update your response parsing — images now arrive nested inside response.candidates[0].content.parts rather than as a flat array. If your app used the number_of_images parameter for batch generation, that parameter no longer exists; you’ll need to loop individual calls instead.

Does this affect Vertex AI too, or just the Gemini API?

Both, on different timelines. Google had already been retiring GA Imagen access on Vertex AI ahead of the Gemini API’s August 17 date. If you’re running Imagen through Vertex AI specifically, check Google Cloud’s own deprecation documentation, since Vertex’s schedule has run separately from the Gemini Developer API’s.

Can I still use Imagen 3 if I’m not ready for Gemini 3.1 Flash Image?

Check the current status directly — Google’s deprecations page has already moved Imagen 3 through earlier retirement stages on its own separate timeline, and shutdown dates on that page are described as the earliest possible dates rather than fixed guarantees. Don’t treat an older Imagen version as a stable fallback without confirming its current status first.

---

Last updated: August 18, 2026. Sources: Google Gemini API deprecations · Google Gemini API pricing · Gemini 3.1 Flash Image (Nano Banana 2) — Google DeepMind · byteiota: Imagen 4 Shutdown Migration Guide · Kingy AI: Imagen 4 API Shutdown Tracker · Google Developer Forums: Nano Banana portrait consistency thread.

Related reading: Gemini 3.1 Pro Review: Google’s Reasoning Breakthrough · Best AI Image Generators in 2026 · The LiteLLM Breach Just Hit 2,500 AI Tool Stacks · Stripe Just Bought OpenRouter for $7 Billion