Google releases Nano Banana image editing model

Background

Editing a picture by describing the change in words — swapping a background, removing an object, restyling a scene — has become one of the standard tests of an image model. It is harder than generating an image from scratch, because the model has to change what the user asked for while leaving everything else intact.

Google framed its August release as part of a continuing line of work rather than a one-off experiment: it said Nano Banana was the latest upgrade to image generation in the Gemini app ^[1].

What happened

On 26 August 2025 Google brought a new image editing model, nicknamed Nano Banana, into the Gemini app ^[1]. The nickname is the kind of informal label that often spreads faster than a version number, and Google used it alongside the product announcement.

The same model was made available to developers under a more conventional name, Gemini 2.5 Flash Image, accessible through the Gemini API and Vertex AI ^[2]. Google priced it at $0.039 per image ^[2]. At launch the developer version was in preview, with a stable release planned for the coming weeks ^[2].

Key details

Google called it the top-rated image editing model in the world on the LMArena image-edit leaderboard ^[1]. LMArena is a public comparison site where models are ranked by blind pairwise votes from users, so a high position there reflects how people judged paired outputs rather than an internal benchmark.

Google also emphasised one particular property of the model: keeping people and pets looking like themselves across edits ^[1]. That is a pointed claim, because a common failure of image editing is that a face or an animal drifts — the person in the edited photo is recognisably not quite the person in the original.

Images created or edited in the app carried both visible and SynthID watermarks ^[1]. SynthID is Google's watermarking technology, which embeds a signal into the pixels themselves so that an AI-generated or AI-edited image can later be identified even when no label is displayed; the visible mark is a separate, plainly noticeable label on top of that.

Why it matters

Two things make a release like this more than a feature update. The first is the price. A per-image figure gives developers a predictable unit cost when they build editing into an app, instead of a metered, hard-to-forecast bill ^[2].

The second is the leaderboard claim. Image editing had become a crowded field, and naming a specific public ranking is a way of staking out a position in it ^[1]. At the same time, the preview-then-stable schedule shows that the developer side was not yet finished at launch ^[2].

Marking the output is the third thread. Pairing a visible mark with SynthID means an edited picture carries an indication both to a person looking at it and to software inspecting it ^[1]. For a tool designed to alter photographs of real people and pets while keeping them recognisable, that is the part of the announcement that concerns everyone outside the app as much as the people using it.