Google Kills Imagen 4: Image API Dies Today, Nano Banana 2 Takes Over

If your application still calls an imagen-4.0-* endpoint this morning, it is already returning errors. Google has executed one of the most consequential API deprecations of the year: as of August 17, 2026, the entire Imagen 4 family on the Gemini Developer API and Vertex AI — imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, and imagen-4.0-fast-generate-001 — is officially dead.

The company announced the deprecation back on June 15, 2026, giving developers roughly a two-month migration window. That window closed today. If you missed it, you are not alone — Firebase’s own migration guide was still surfacing in developer forums this weekend as people scrambled to patch production pipelines over the holiday lull.

What exactly is shutting down

Three API endpoints form the kill list:

• imagen-4.0-generate-001 — the standard tier, Google’s recommended default for most text-to-image workloads
• imagen-4.0-ultra-generate-001 — the highest-quality tier, used for photorealistic hero assets
• imagen-4.0-fast-generate-001 — the low-latency tier, popular in interactive and high-volume applications

Google’s replacement path is unambiguous: migrate to the Gemini 3.x Image family, colloquially known as the “Nano Banana” models. The Firebase migration guide maps every deprecated Imagen endpoint to a specific successor:

Deprecated endpoint	Replacement
imagen-4.0-fast-generate-001	gemini-3.1-flash-image (with thinking level)
imagen-4.0-generate-001	gemini-3.1-flash-image (with thinking level)
image-4.0-ultra-generate-001	gemini-3.1-flash-image

All three tiers collapse into a single model: Gemini 3.1 Flash Image, or “Nano Banana 2,” which Google has billed since its February launch as its best image generation and editing model with strong price-performance via the Gemini API.

The migration is not a drop-in swap

The trap in this transition is that Nano Banana 2 is architecturally a different animal from Imagen. Imagen was a dedicated text-to-image diffusion service with its own request schema. Gemini 3.1 Flash Image is a natively multimodal Gemini model, which means request parameters, response parsing, and even safety configuration work differently.

Three changes bite hardest in practice:

1. Request schema. Generation and editing calls move to the standard Gemini generateContent surface. Fields like numberOfImages and Imagen-specific aspect-ratio parameters don’t carry over verbatim.
2. Safety settings. Imagen’s dedicated safety filters are replaced by Gemini’s SafetySetting objects — code that relied on filter-specific error codes needs rework.
3. Response format. Images arrive as inline parts of a multimodal response rather than a discrete image payload, so parsing logic in existing pipelines must be rewritten.

None of this is hard, but none of it is a find-and-replace either. Teams that treated the two-month window as generous discovered the real migration cost sits in testing visual output consistency, not in swapping a model string.

The cost question nobody can ignore

Here is where the migration gets politically uncomfortable. Independent analysis from TokenCost pegged the cost of replacing Imagen 4 workloads with Gemini 3.1 Flash Image at roughly 3.36x the previous spend for comparable usage patterns — a consequence of moving from per-image diffusion pricing to token-based multimodal pricing, where a single high-resolution generation can consume substantially more tokens than an Imagen call did.

Google’s official pricing page currently lists Gemini image generation in the vicinity of $0.151 per image at the high end, with standard-resolution generation around $0.0336 per image under standard tiers. Depending on which Imagen tier you were on and your resolution mix, your bill could move in either direction — but the developers hit hardest are precisely the high-volume Fast-tier users who chose Imagen specifically because it was cheap.

The irony is thick. One of the major selling points of the Nano Banana family was that it made image generation “democratic” and accessible. For pipeline operators, the shutdown converts that story into a cost increase and a forced refactor at the same time.

A second deadline is already looming

Developers respired by today’s shutdown should note that Google has already scheduled the next one: Gemini 2.5 Pro and Flash models shut down on October 16, 2026. That includes the older gemini-2.5-flash-image endpoint that some rushed migrations mistakenly target — Google’s own guidance says to jump directly to gemini-3.7-flash or gemini-3.1-flash-image to avoid chaining migrations. The deprecation treadmill is accelerating, and betting on any single endpoint has become an operational risk in itself.

The bigger picture: consolidation, not expansion

The Imagen shutdown is the clearest signal yet that Google is consolidating its entire generative media stack onto the Gemini chassis. Imagen as a brand existed because diffusion-era Google needed a dedicated image model family. In the Gemini era, image generation is just another modality of a unified model — and maintaining a parallel API surface for it no longer makes economic sense for Google.

For developers, the lesson is structural: image-generation APIs attached to “legacy” model families are living on borrowed time across the industry, not just at Google. The pragmatic posture is to abstract image generation behind your own interface layer, keep provider swap-tests in CI, and treat any provider’s deprecation calendar as a first-class input to architecture decisions.

Today the Imagen era ends not with a bang but with a 404. If your pipeline is still pointing at imagen-4.0-*, the fix window is now — and the next deadline, October 16, is only eight weeks away.