In the middle of the night, Google attacked OpenAI!

  It's the first time in the history of the world.

Google has just released Nano Banana 2.1 in the middle of the night!

A Flash-sized small model has taken its flagship Nano Banana Pro down.

I'm going to focus on this:

• Modified: Support for local editing of printed editions, where the circle has been changed, elsewhere is still moving
• Consistency: characters and products have changed several rounds without changing faces, with a maximum of 14 reference charts.
• Image quality: more like a live shot, up to 4K.
• Pricing: Half of Nano Banana 2 and one K for $0.034.
• List: 4th edition of Arena Graphics, ahead of only three OpenAI GPT Images
• Top: Gemini App, search AI mode, AI Studio, Flow and more are available today

First, Nano Banana 2.1 is also a member of the Gemini 3 series of models, based on Gemini 3.6 Flash, with a knowledge deadline of March 2026.

Nano Banana 2.1 is also known as Gemini 3.6 Flash Image.

According to official benchmarks, Nano Banana 2.1 is far ahead in terms of visual design, editing, and consistency.

While there is still room for improvement in small-text rendering, 3D reasoning, and factuality, this is undoubtedly a pretty great update for users who need image generation and editing services:

• Create and edit images with professional level of accuracy and control, and support multiple rapid iterations
• Create clear text for posters and complex charts
• Growing knowledge of the real world in context
• Localized translations across multiple languages

The results of the blind vote by netizens also prove this.

At Arena Arena, Nano Banana 2.1 took the 4th edition, the 5th edition, and the 6th edition.

Especially on the multi-graphic editing list, the only remaining three GPT Image from OpenAI are in front of it.

In comparison, the previous generation Nano Banana 2 ranked only 7th, 11th and 14th in these three charts.

Scroll to the left and look.

Given that Nano Banana 2 is only half the price, it looks like Google is going to have a price fight.

The first edition of the book was edited with the help of an editor who was familiar with the PS.

It has always been difficult to put AI in the living room.

If you tell a model to "turn the bottom-left corner cup into red", it has to guess which cup you're talking about, and then reproduce the whole picture again, which can't be saved anywhere else, with good luck.

It's time to make a circle. Change your clothes, change your badge, turn off the background.

Ibexdream gives it three questions: generate, modify, and change style.

The first question is to draw a piece of paper. The picture shows an old philosopher with his hair turned backwards into a maze.

The second title is in this chart. The number was changed from "No 1098471" to "NB-21098471", the lowercase letters on both sides were changed, the old philosopher had more champagne in his hands as a phone call, and a ribbon with the inscription "IN BANANA WE TRUST" at the bottom.

When they put the two together, the labyrinth, the robe and the beard remained the same.

The third issue changed the style, and the whole of the banknotes became an oil painting.

In the rating of the model card, the modified version 2.1 received 1,049 points, 84 points higher than Nano Banana 2 and 122 points higher than Pro.

In fact, even if you don't think about it, 2.1 pushes Pro on every single one of them.

Another upgrade is subject consistency, which means don't change your face.

People who make e-commerce photos, make brand posters, draw cartoons should be aware that the same model changes 10 sets of clothes to a third one, and their faces start to shrink.

The highest score was this one. Multi-character consistency scored 1106 points, 128 points higher than Nano Banana 2.

The developer can fill up a maximum of 10 objects with reference pictures, four characters with reference pictures, and make it look like a picture.

The netizens BG-VC tried this. He gave a model photo, plus a reference picture of five individual items of clothing, hemorrhoids, bags, shoes, earrings, to make it look like six different scenes, six of which were generated simultaneously.

Walking, sitting, leaning on the counter, changing six postures, faces, clothes, bags and earrings didn't work out, and taking a set of e-commerce maps was enough.

The picture itself is more like it was taken. In the official samples, the green girl on the stairs outside the house, hung with beetles and water pearls, is an AI painting.

In addition, when making an infographic, it can first go to Google's web page search and image search to find information and then click on the pen.

In Google's auto-review, the accuracy of 2.1 is 0.521, Nano Banana 2 is only 0.179, and Pro is only 0.265.

In other words, let it paint a copy, and there will be fewer cases of mistakes in the content.

It's $30 per million token, which is half of Nano Banana 2.

1K decreased from $0.067 to $0.034, and 4K decreased from $0.151 to $0.076.

This price is the same as the Nano Banana 2 Lite, which was the main dish in July.

However, other projects failed.

Input earned from $0.5 to $1.5 per million of tokens, three times the original output. Text and thought earned from $3 to $7.5 as well.

It's like Google moving money from "painting" to "thinking".

Currently, developers can now select it in AI Studio, the model is called gemini-nano-banana-2.1.

The previous Nano Banana 2 API will be shut down on October 29.

Ordinary users don't have to wait.

Gemini App, search AI models, Flow, design tools, Stitch, Google Ads and the Gemini enterprise platform are online today.

In the Google App, the banana icon under the search box can be played.

The model went live in less than an hour, and someone did the measurements on X.

Mark Kretschmann gave it a "torture test", which suggested that one piece of paper was simultaneously filled with people, dirt, clear text, reflections of rain, and then looked at them together, as if they were not a real photo.

Luis Catacora had it paint a menu board in a teahouse, the same "Jelly Green Tea", written in four letters in English, Japanese, Arabic, and Hindi, one at a time.

Lad wrote a specially selected AI soft rib prompter with a fine-grained tattoo, three necklaces, a row of earrings, a handkerchief, and a pair of eye-catching colored lenses, and then handed the prompter to Nano Banana 2.1 and GPT Image 2.5 respectively.

Below is a compilation of Nano Banana 2.1 on the left and GPT Image 2.5 on the right.

Nano Banana 2.1 painted the entire tattoo in his designated position, including necklaces, earrings, and gloves, which means the lens has forgotten its color and the face is red.

Pankaj Kumar's opponent was GPT Image 2. He had the two models draw the same scene, placing a watch on the desk, one each.

On the left is Nano Banana 2.1, with natural light, like hand-painted. On the right is GPT Image 2, with the counterlight turned on and the dashboard switched to a vacuum motion, a taste of refined advertising graphics.

The API platform AI/ML API prepares five suggestive words, each for Nano Banana 2.1 and GPT Images 2.5.

Nano Banana 2.1 takes 11 seconds and costs $0.044, while GPT Images 2.5 takes 50 seconds and costs $0.069.

Hugo Plat came up with four space puzzles, spinning galaxies, supermassive black holes, and galaxies, adding a picture of the entire solar system, which was also given to Nano Banana 2.1 and GPT Image 2.5.

The Nano Banana 2.1 cost $0.178 and the GPT Image 2.5 cost $0.284.

The same black hole theme, Nano Banana 2.1 on the left, with controls, and GPT Image 2.5 on the right, full of clouds and flames.

This piece in Chinese, a well-known blogger attributed to a round of testing, concluded that the words were clearer, less misspellings, the layout cleaner, and without the fragility of GPT. He also felt that, compared to GPT, its biggest advantage was that it could directly output 4K.

Scroll to the left and look.

Of course, it's not without its flaws. The little English line next to "Legacy and Newcomers" is blurred.

In April, GPT Image 2 came online, leading Nano Banana 2 by 240 points on the same Arena's chart.

In September, OpenAI released another version of Images 2.5, which focused on changing the designated area only.

In less than a month, Google's 2.1 bet on both.

It's not hard to guess why the two front and back scrolls have changed. E-commerce, advertising and brand designs have changed these lines, one of the drawings has been changed to dozens of editions, and the tenth edition has not collapsed, which is much more important than the first edition surprise.

So this time, Google is directly plugging 2.1 into Google Ads and enterprise platforms, and that's the group of customers who are changing every day.

Speaking of which, when is the next generation of Nano Banana Pro going to come out?