Google on Tuesday, Oct. 6, released Nano Banana 2.1, the newest version of its AI that turns a typed sentence into a picture or edits a photo you already have. The model is rolling out across the Gemini app, Google Search's AI Mode, Google Ads, and developer tools including Google AI Studio, Flow, and Stitch — a spread that puts a single image model behind consumer chat, search, and advertising surfaces at the same time.

Meet Nano Banana 2.1, our latest image generation and editing model 🍌 This upgraded version outperforms our previous models across the board, with notable leaps in visual design, mask-based editing, and subject consistency to help you create more natural-looking images. pic.twitter.com/d858DrzrF3

— Google (@Google) October 6, 2026

According to Google's model card, the update concentrates on three areas. Visual design has been improved to help produce more natural-looking images; mask-based editing is sharper, so users can mark one region of an image and change only that area; and subject consistency is stronger, meaning a person or object still looks like itself after several edits. Together, those three target iterative editing — the kind of workflow where an image passes through multiple rounds of changes and has to stay coherent from one edit to the next.

Google's own tests put the new model ahead of both Nano Banana 2 and Nano Banana Pro. On overall preference in text-to-image evaluations, Nano Banana 2.1 scored 1,050 ELO points, compared with 990 for Nano Banana 2 and 935 for Nano Banana Pro. In the ELO system, higher scores are better, and the scale has no ceiling, making the ratings a purely comparative measure of preference between outputs rather than performance against a fixed technical standard. Google also reported an infographic factuality score of 0.521 versus 0.179.

On paper, the model handles up to 14 reference images at once, tracking as many as four characters and ten objects. It outputs images at up to 4K resolution — roughly TV-level sharpness — in formats as wide as 8:1. Developers can also set how long the model "thinks" before drawing, from minimal to high, and enable grounding in Google Search and Google Image Search. In plain terms, it can look things up before it draws them.

Pricing is a central part of the release. Through Google's developer API — the pipeline that lets other companies' apps use the model — a standard 1K image now costs $0.0336, about half of Nano Banana 2's $0.067, according to Google's pricing documentation. A 4K image runs $0.0756, down from $0.151. Batch jobs, in which large piles of images are processed in bulk rather than generated instantly, receive another 50% discount on top of those rates. At the standard 1K price, 1,000 images now cost about $33.60 through the API, down from $67 on Nano Banana 2, making high-volume image generation on the platform roughly half as expensive as before. Because API billing is charged per image, that rate is the number developers multiply across production workloads — which is why the cut matters most for products outside Google's own apps that generate pictures at scale.

The launch continues a fast cadence of upgrades. The original Nano Banana made Gemini a hit: in September 2025, its trick of turning selfies into collectible figurines went viral, pushing Gemini to No. 1 on both major app stores and dethroning ChatGPT after nearly three years at the top. Alphabet's market value topped $3 trillion in the same stretch. That precedent has made image generation one of Gemini's most visible consumer-facing capabilities.

Version 2 arrived in February, built on Gemini 3.1 Flash Image. Its key feature was checking Google Search before drawing real-world things, giving pictures of actual events or people a better chance of being accurate.

Independent benchmarking of Nano Banana 2.1 was not available at launch; the published comparisons came entirely from Google's internal testing. That distinction is the one to watch from here: independent evaluations, once published, are the outside check on whether the preference and factuality gains hold up beyond Google's own numbers.