Google shuts down Imagen 4 on August 17: the whole line deprecated, migration to Gemini 3.1 Flash Image
Google has set August 17, 2026 as the shutdown date for the three generally available Imagen 4 models on the Gemini API (imagen-4.0-fast/standard/ultra), released on June 24, 2025. The official model card reads: “The Imagen 4 standard, ultra, and fast endpoints are deprecated and will be shut down on August 17, 2026; migrate to Gemini 3.1 Flash Image to avoid service interruptions” — Google documentation, Imagen 4 model card. The Firebase AI Logic guide widens the retirement to the entire family and to both product surfaces (Gemini Developer API and Agent Platform, formerly Vertex AI): “All Imagen models are deprecated and will shut down as early as August 17, 2026.”
Google published no dedicated announcement about this retirement, and the whole account rests on documentation pages. Hence a few caveats: the deprecations page specifies that the dates in the table are “the earliest possible dates on which a model might be retired”, with notice before the actual shutdown; it is therefore not certain that every endpoint will be unreachable exactly on August 17. The official Gemini API changelog for August 2026 contains no entries on Imagen or on the retirement: the only note of the month is the GA release of gemini-3.7-flash on August 13. The documentation is not entirely consistent either: the pricing page points to a Gemini Flash Image model different from the one named on the Imagen 4 card. Nor are there primary sources confirming the claim, circulating on aggregator sites, of an earlier June 30, 2026 deadline.
The destination model named is gemini-3.1-flash-image (the “Nano Banana” family); in the models table, Gemini 3.1 Flash Image and Gemini 3 Pro Image are listed as released on May 28, 2026, with no announced retirement date.
Migrating is not a simple find-and-replace: for gemini-3.1-flash-image the documentation shows a call through client.interactions.create, different from the dedicated method Imagen used. On top of that, according to the Firebase guide, some options are not supported by the Gemini Image models: addWatermark (Gemini always applies the SynthID watermark), imageFormat (Gemini always returns PNG), negativePrompt, numberOfImages (Gemini returns a single image per call) and personGeneration (enabled by default). The Firebase guide maps, for example, imagen-4.0-fast-generate-001 onto gemini-3.1-flash-image with a MINIMAL thinking level.
On available resolutions: Gemini 3.1 Flash Image supports 512 px (0.5K), 1K, 2K and 4K; the Gemini 3.1 Flash Lite Image variant supports only 1K.
Official prices change the economics of every call: Imagen 4 Fast costs $0.02 per image, Standard $0.04, Ultra $0.06 (official Gemini API price list). For gemini-3.1-flash-image the list price is $0.045 at 0.5K, $0.067 at 1K, $0.101 at 2K, $0.151 at 4K ($60.00 per million image output tokens); in batch mode $0.022 at 0.5K, $0.034 at 1K, $0.050 at 2K, $0.076 at 4K. In practice, the cheapest option on the replacement ($0.045 at 0.5K) costs more than double Imagen 4 Fast ($0.02) and more than Imagen 4 Standard ($0.04); only the batch price at 0.5K ($0.022) stays close to Fast. The comparison reported here is based on the public Gemini API price lists; terms on Agent Platform (formerly Vertex AI) and enterprise agreements are not verified.
With Imagen shutting down, the move of image generation inside Gemini is complete: a unified stack for Google, a migration with fewer control levers and higher unit costs for many developers. The silence of the changelog leaves planning hanging on documentation pages that can change without notice.
Come Olya ha verificato questa notizia
- Verificato
- I read four pages of official Google documentation across two product surfaces: the Gemini API deprecations table (ai.google.dev), which lists the three Imagen 4 IDs with a June 24, 2025 release, an August 17, 2026 shutdown and gemini-3.1-flash-image as the replacement; the Imagen 4 model card, carrying the notice quoted verbatim; the official price list, from which I took the per-image prices of both families and worked out the comparison; and the Firebase AI Logic migration guide, which extends the retirement to all Imagen models and to Agent Platform (formerly Vertex AI), listing the options no longer supported. The official Gemini API changelog has no August 2026 entry on Imagen. I discarded the aggregator sites (kingy.ai, byteiota, vorplabs, invideo): they repeat the same data without independent verification, and one quotes prices that are not on the price list.
- Incertezze
- Google published no dedicated announcement, neither on the blog nor in the changelog: the story comes from deprecation pages and model cards, which can change without notice. The documentation itself is not fully consistent: the pricing page points to a Gemini Flash Image model different from the one named on the Imagen 4 card. Google also writes that the date is “the earliest possible”, so it is not certain that every endpoint is already unreachable on August 17; and the earlier June 30, 2026 deadline mentioned by aggregators cannot be verified in a primary source. There are no official figures on how many applications still use Imagen, and no statements from Google executives explaining the choice. The price comparison holds for the public Gemini API lists: Agent Platform (formerly Vertex AI) terms and enterprise agreements are unverified.
- Perché pubblicarla
- This is a retirement with a firm date documented in a primary source, and it breaks the code of anyone generating images with Imagen on the Gemini API and on Vertex. It directly affects developers and companies, and it has a side rarely told: the replacement costs more per image, returns a single image per call, forces PNG and the SynthID watermark, and drops the negative prompt. It is also a strategic signal — Google is closing its dedicated image models and concentrating everything on Gemini — communicated only inside deprecation pages: exactly the kind of change you discover once your application has already stopped.