Google Nano Banana 2.1: Improved image model for half the price

Google revamped its image model with Nano Banana 2 in February, and another update is already here. Nano Banana 2.1 is based on Gemini 3.6 Flash and, according to Google, outperforms its predecessors across the board, particularly in visual design, editing specific parts of an image, and maintaining the appearance of people and objects across multiple editing steps. Google is gradually rolling out the model to the Gemini app, AI Mode in Google Search, Google Ads, and developer tools including AI Studio, Flow and Stitch.

For now, Google's own figures are the only indication of how much has improved. In comparison tests with human raters, Nano Banana 2.1 scores 1,050 Elo points for overall preference with Thinking enabled, compared with 990 for Nano Banana 2 and 935 for the more expensive Nano Banana Pro. The gap is particularly wide for infographics: the new model scores 0.521 for factual accuracy, versus just 0.179 for Nano Banana 2. Independent tests are not yet available.

Select an area instead of regenerating the image

Mask-based editing is likely to be the most useful improvement in everyday use: users select an area of an image, and only that area changes. Signs of this feature had already appeared in the Gemini app in March. Nano Banana 2.1 also handles up to 14 reference images at once, including up to four people and ten objects whose appearance should be preserved. It supports output resolutions up to 4K and aspect ratios up to 8:1. When needed, the model consults Google Search before depicting real-world subjects.

Lower image prices, higher input costs

Google is roughly halving the price per image for developers: through the Gemini API, a 1K image costs about 3.4 cents instead of 6.7 cents, while a 4K image costs about 7.6 cents instead of 15.1 cents. Batch jobs cost 50% less. The calculation is not quite that simple, however, because input costs are rising: Google now charges $1.50 per million tokens for text, image and video inputs, up from $0.50. Developers working with many reference images will therefore save less than the halved image prices might suggest.

Does that leave a role for Nano Banana Pro? For many tasks, probably not, provided Google's figures hold up in practice. The new model is still imperfect, as Google acknowledges in its model card: small text often remains blurry at 1K resolution, people do not always look exactly like their reference images, and the model occasionally confuses left and right. Nano Banana 2.1 is therefore a solid update rather than a leap on the scale of the original version, with price as its strongest argument.