Google launches Nano Banana 2.1: better images at half the price

Google has released Nano Banana 2.1, the newest version of the image model that turned the Gemini app into a viral hit last year. The update arrived on 6 October 2026 and comes with two messages at once: the pictures are supposed to be better across the board, and the price for developers drops to roughly half of what Nano Banana 2 cost. A standard 1K image now costs 3.36 US cents through the API instead of 6.70 cents. A 4K image falls from 15.10 to 7.56 cents.

That combination matters more than it sounds. Image generation is no longer a toy feature that people try once for a funny selfie. It sits inside ad tools, design apps, presentation builders and ecommerce pipelines that create thousands of pictures every day. When the unit price halves while quality goes up, the economics of all those products change overnight.

This article explains what Nano Banana 2.1 is, what Google says it does better, what it costs, where you can use it, and where you should stay sceptical. The figures come from reporting by The Decoder and Decrypt, both based on Google's launch material and model card. All benchmark numbers are Google's own tests. Independent results were not available at launch.

A quick history of Nano Banana

The name sounds like a joke, and it started as one. The original Nano Banana was Google's codename for Gemini 2.5 Flash Image. In September 2025 its trick of turning ordinary selfies into collectible figurine style images spread across social media and pushed the Gemini app to the top of both major app stores, ahead of ChatGPT for the first time in almost three years. Google kept the playful name as an official brand.

Version 2 followed in February 2026, built on Gemini 3.1 Flash Image. Its headline feature was grounding: before drawing real world things, it could check Google Search, so a picture of an actual building, event or product had a better chance of being correct. Alongside it sits Nano Banana Pro, built on Gemini 3 Pro Image, which is slower and more expensive but often produces the most realistic results.

Nano Banana 2.1 replaces Nano Banana 2. According to the model card it is based on Gemini 3.6 Flash, so it inherits a newer and stronger language model under the hood while keeping Flash tier speed. Google describes it as the more efficient counterpart to Pro rather than a replacement for it.

The price cut in numbers

The most concrete part of the launch is the price list. Google charges image output by tokens, which works out to a fixed price per picture depending on resolution. The standard output rate is 30 US dollars per million tokens.

At 1K resolution, Nano Banana 2.1 costs 3.36 cents per image. That is exactly half of the 6.70 cents for Nano Banana 2, and a quarter of the 13.40 cents Google charges for Nano Banana Pro. It also matches the price of the cheaper Nano Banana 2 Lite, which means developers no longer have to choose between the small model and the good model at that resolution.

To make this tangible: a thousand standard images now cost about 33.60 US dollars instead of 67 dollars. A shop that renders product photos in ten variations for every new item, or an ad platform that generates several creatives per campaign, feels that difference immediately in its monthly bill.

The cut holds at higher resolutions too. A 2K image costs 5.04 cents, down from 10.10. A 4K image costs 7.56 cents, down from 15.10, while Pro sits at 24 cents for the same size. On top of that, batch jobs, where images are processed in bulk instead of instantly, get another 50 percent off. For workloads that do not need an answer within seconds, the effective price for a 1K picture drops below two cents.

What Google says is better

Google claims Nano Banana 2.1 improves on its previous models across the board. The company highlights a handful of areas in particular.

• Visual design. Layouts, typography and composition are meant to look more deliberate and less like a random AI picture.
• Text rendering. Words inside images, on signs, packaging or posters, should be spelled correctly more often.
• Character consistency. A person or object should still look like itself after several rounds of edits in the same conversation.
• Mask based editing. You mark one region of a picture and only that region changes, while the rest stays untouched.
• Panoramas and infographics. Wide formats up to 8:1 and fact based graphics are explicitly named as strengths.

On the technical side, the model accepts up to 14 reference images in a single request and can keep up to four characters and ten objects consistent across them. That is useful for storyboards, comics, product catalogues or any series where the same person or item has to appear again and again. Output is available at 1K, 2K and 4K.

Developers also get control over how much the model thinks before it draws. There are minimal, medium and high thinking levels. More thinking means more time and usually better results, especially for complex prompts. And like its predecessor, the model can ground its work in Google Search and Google Image Search, which in plain words means it can look something up before drawing it.

The benchmarks, and how to read them

Google published preference scores from its own evaluations. These are Elo style ratings, the same system used in chess and on public AI leaderboards: people compare two images for the same prompt and pick the better one, and the scores reflect how often each model wins. Higher is better, but the numbers only make sense relative to each other.

In overall text to image preference, Nano Banana 2.1 with thinking scored 1050, without thinking 1015, Nano Banana 2 scored 990 and Nano Banana Pro only 935. The editing categories show a similar picture. The largest gap appears in multi character consistency, where 2.1 with thinking reaches 1106 against 978 for Nano Banana 2 and 1011 for Pro.

The most striking number is not an Elo score at all. For infographics, Google measured how often the generated graphic was factually correct. Nano Banana 2.1 with thinking reached 0.521, compared with 0.179 for Nano Banana 2 and 0.265 for Pro.

That is almost three times the accuracy of the previous Flash model. It also shows clearly what thinking mode buys you: without it, the score drops to 0.328. At the same time, a value of 0.521 means that roughly every second infographic still contains an error. Anyone planning to use AI generated charts in reports or on social media should keep checking every number by hand.

There is an important caveat to all of this. These are vendor benchmarks. Nano Banana 2 also beat Pro in Google's earlier tests, yet many users found that Pro still produced noticeably better images in practice. The Decoder ran its own standard prompt through 2.1 and concluded that Pro's result still looked more realistic and natural, especially in colours and proportions. In that test, 2.1 struggled with scale and drew a horse that looked more like a pony. So the fair summary is: clearly better than Nano Banana 2, cheaper, faster, but not automatically better than Pro for photorealism.

Where you can use it

Nano Banana 2.1 is rolling out across Google's products at once. For regular users, that means the Gemini app and AI Mode in Google Search. For creators and designers, it shows up in Flow, Google's video and image tool, and in Stitch, the app design tool. Advertisers get it in Google Ads. Developers can call it through Google AI Studio and the Gemini API, and businesses through the Gemini Enterprise Platform.

There is one deadline developers should note. Google will shut down the previous model, listed as gemini 3.1 flash image, on 29 October 2026. Any app that still calls the old model name needs to switch before then. Given the lower price, most teams will not mind, but it is worth testing prompts in advance, because a new model can change the look of outputs that users have grown used to.

Why the price war matters

The price cut does not happen in a vacuum. Image generation has become one of the most visible fields in the race between AI labs. OpenAI offers image generation inside ChatGPT and through its API, Mistral and several open model providers compete on price, and specialised companies like Midjourney and Black Forest Labs fight for creative professionals.

By halving its price while claiming better quality, Google puts pressure on all of them. The company can afford it because it runs its own chips and its own data centres, and because images feed directly into products where it makes money, above all advertising. A cheaper, better image model makes Google Ads more attractive for small businesses that cannot pay a designer for every campaign.

For developers, the effect is simple. Features that were too expensive at seven cents per image may suddenly make sense at three cents, or below two cents in batch. Think of personalised product shots for every visitor, automatic thumbnails for every article, or illustrated summaries of long documents. Many of these ideas fail not because the technology is missing, but because the unit cost does not add up at scale.

What to keep in mind

Better and cheaper image models also raise familiar questions. Stronger character consistency and accurate text rendering make it easier to produce convincing fakes, from fake product photos to manipulated screenshots. Google marks images from its models with SynthID, its invisible watermark, and adds visible indicators in some products, but such marks only help when platforms and users actually check them.

For businesses there is also a practical point: vendor benchmarks are a starting signal, not a guarantee. Before moving a production workload, run your own prompts through Nano Banana 2.1 and compare the results with what you use today. Look closely at faces, hands, small text and proportions, because those are the areas where models still differ most.

Bottom line

Nano Banana 2.1 is not a revolution, but it is a very strong update. Google claims gains in design, text, editing and consistency, and the infographic accuracy jump is the most impressive number in the launch, even if half of all infographics still need fixing. The bigger story is the price: half of Nano Banana 2, a quarter of Pro, and even less in batch mode.

If you build products on image generation, now is a good moment to test it, and you have a deadline: the old model disappears on 29 October. If you simply use the Gemini app, you should notice better pictures soon without doing anything. And if you need the most realistic photos possible, Nano Banana Pro is likely still the better choice, for now.

Sources: The Decoder: Google's new image model Nano Banana 2.1 generates better images for less money and Decrypt: Google launches Nano Banana 2.1, both published 6 October 2026.