Google launches Nano Banana Pro

What Google announced

On 20 November 2025 Google introduced Nano Banana Pro, an image-generation model built on Gemini 3 Pro.^[1] The announcement put most of its weight on a single capability: getting the text inside a picture to come out right. Google described the model as the best one for creating images with correctly rendered and legible text directly in the image.^[1]

The features Google highlighted

Three things were singled out. First, legible text in multiple languages, meaning the model is meant to spell words correctly in more than one writing system rather than merely produce letter-like shapes.^[1] Second, Search-grounded infographics: pictures of the explanatory-chart kind whose content is drawn from search results rather than from the model's memory alone.^[1] Third, output resolution — images could be generated at 2K and 4K.^[1]

Google also gave the model a compositing ability. It can blend up to 14 input images, and it can keep the likeness of up to 5 people.^[1] Many reference images plus a handful of recurring faces is exactly what a user needs when assembling a group scene or a product mock-up out of separate sources.

Who could use it

Access was tiered. Free-tier users of the Gemini app received limited free quotas, while paid tiers came with higher quotas.^[1] In other words, the model shipped inside the existing app rather than as a separate product, with usage metered by plan.

Verification on the same day

Google added SynthID image verification inside the Gemini app on the same day.^[1] Verification means a user can check whether a picture carries Google's label for AI-generated content, so an image seen in a chat or a feed can be tested rather than guessed at.

Why text is the hard part

Image models work on a compressed internal representation of a picture, not on a sequence of characters. Nothing in that representation is anchored to the shapes of letters, so earlier models reproduced writing as texture: convincing at a glance, misspelled on inspection. Making text legible means holding letterforms stable across a whole caption or a row of chart labels, and keeping them correct in languages other than English. That is why a launch that leads with correct spelling and readable labels is a substantive claim about the model rather than a styling detail: charts, posters, diagrams, slides and menus all depend on words that can actually be read.^[1]

Where the release sits

Taken together — a model built on Gemini 3 Pro, higher-resolution output, multi-image input, and a verification tool arriving the same day — the release treats image generation as a way of making informational pictures, not only attractive ones. For anyone judging whether an image model is usable for real work, the first question is often whether the line of text on the picture reads correctly.