The newest version of Google’s AI image model is here, thanks to Gemini 3. Whether you choose to call it by its official name (Gemini 3 Pro Image) or its better-known nickname (nano banana pro), this model uses Gemini’s “deep thinking” feature to create better content.
Nano banana pro promises to be more capable, specifically with generating legible text, using up to 14 reference images and creating images in resolutions up to 4K.
You can upload up to 14 separate reference images to nano banana pro. Notice how each of the characters in the prompt (left) appear unchanged in the final image (right).
Nano banana pro is available now, rolling out globally in the Gemini app. In Gemini, select “Create images” (with a 🍌 emoji) and switch your model to the Thinking model (located on the right side of the prompt window).
You can use nano banana pro for free, but you will hit a generation limit faster than paying Google AI subscribers. Those plans start at $20 per month. You’ll be automatically switched back to the original model once you’ve used up your credits. The model is also available in the Gemini API, Google AI Studio and Flow.
Inside the new nano banana model
AI image generators have notoriously struggled to create clear text; mangled words have been clear signs that an image was created with AI. OpenAI’s GPT-Image-1 was one of the first models to make significant progress in fixing this, but CNET’s testing showed it wasn’t able to reliably create readable text.
Google says nano banana pro is supposed to be better at this — so good that you can use the new model to create infographics, relying on Gemini’s integration with Google Search (and Search’s AI Mode for paying Pro users) to aid the process.
An example of an infographic made with nano banana pro.
You might’ve noticed that none of the images made with nano banana pro have Google’s diamond-shaped watermark, which usually denotes that it was made with AI. That’s because Google is introducing a new way for its highest-paying subscribers (those paying $250 per month for the Ultra plan) to remove those watermarks. Even though it’s restricted to that top-tier plan, it’s already extremely difficult to tell if something was created with AI. Fewer watermarks won’t help. Google says Gemini will be able to identify any images made with its AI because of its SynthID, an invisible watermark attached to its AI-generated content. So in theory, you can upload an image to Gemini and ask whether it’s real or made with Google’s AI. We’ll have to wait and see how well it actually works.
Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.
The original nano banana model, Gemini 2.5 Flash Image, was released this summer. Early testers thought the capabilities were, well, bananas. Throw in a series of fruit-themed teasers from Google executives, and the model quickly became best known by its nickname, nano banana. Its popularity soared in the following days.
One of the things fans liked about the original nano banana was how well it maintained character consistency. Meaning nano banana was able to edit your existing photos without distorting the characters or people in them. CNET’s hands-on testing found this was true, but there were still obvious AI slip-ups. The new model might eliminate some of those errors.
It’s been a big week for Google as it dropped a major update to its AI system, Gemini 3. The new models are rolling out now, aiming to give Gemini users a smarter, less sycophantic experience.
For more, check out how to write the best AI image prompt and the best AI image generators.
