Google's nano banana is officially launched: The cost per image is less than 0.3 yuan, 95% cheaper than OpenAI.

Last night, the mysterious and powerful image generation and editing model, nano banana, finally revealed its true self. As expected, it comes from Google and has been given an official but uninteresting name: gemini-2.5-flash-image-preview.

According to the introduction, this model has "SOTA image generation and editing capabilities, amazing character consistency, and lightning-fast speed."

Judging from its name, it can be guessed that Google should also have a non-flash gemini-2.5-image model - its performance should be more powerful, but the speed will be slower.

Currently, gemini-2.5-flash-image-preview is already available for preview in Google AI Studio and Gemini API. Users can try it for free.

As can be seen, gemini-2.5-flash-image-preview supports a 32k context, provides temperature (which can control the model's creativity), and some advanced settings.

However, unfortunately, this model does not yet support image generation and editing for Chinese input, but will give a text response instead.

In addition, in Gemini, users can also use this model by simply selecting 2.5 Flash and using appropriate prompts.

In terms of price, the price for input/output text of gemini-2.5-flash-image-preview is $0.3/$2.5, and the price for input/output images is $0.3/$30. The knowledge cutoff time is June 2025.

Approximately calculated, the cost of each image generated by this model is about $0.039 (about 0.28 yuan), which is much lower than the image generation cost of OpenAI.

In terms of specific functions (especially image editing), Google's official blog introduced that they pay particular attention to maintaining the consistency of the character image between different pictures.

"We know that when you edit yourself or someone you're familiar with, even the slightest difference can be jarring - the 'almost but not quite' effect just doesn't feel right. That's why our latest update specifically addresses this, ensuring that your friends, family, and even your pets always look like themselves, whether they're sporting a 60s beehive haircut or dressing a Chihuahua in a tutu."

You just need to give Gemini a photo and tell it what you want to modify, and you can add a unique personal style. This model can help you put yourself and your pet in the same photo, change the room background to the effect of new wallpaper, or take you to any place in the world you can imagine - while still keeping "you as you are." After completion, you can even upload the edited photo to Gemini again and turn the new image into an interesting video.

Google also shared some examples of how to play.

Change clothes or scenes: Upload a photo of a person or a pet, and the model will keep their appearance consistent in any new scene. You can try different clothes, occupations, and even see what you would look like in another era - but still be yourself.

Google even specifically built a demo template application to show what you would look like in different eras.

Address: https://aistudio.google.com/apps/bundled/past_forward

Composite photos: You can now upload multiple photos and merge them into a brand - new scene. For example, composite a photo of you and your dog on a basketball court to generate a perfect group photo.

Multi - round editing: You can continuously modify the images generated by Gemini. For example, start with an empty room, first paint the walls, and then add bookshelves, furniture, or a tea table. Gemini will assist you all the way, only changing the parts you specify while keeping the rest intact.

Mixed design: Apply the style of one image to the objects in another image. For example, apply the color and texture of petals to a pair of rain boots, or design a dress with the pattern of butterfly wings.

Native world knowledge: This model can also utilize Gemini's world knowledge to unlock brand - new application scenarios. To demonstrate this, Google built a template application in Google AI Studio, which can turn a simple canvas into an interactive educational tutor.

Address: https://aistudio.google.com/apps/bundled/codrawing

In addition, Google also mentioned that all pictures generated or edited in the Gemini application will have a visible watermark and Google's invisible SynthID digital watermark to clearly identify that they are AI - generated.

As soon as this model was launched, there was a wave of testing enthusiasm. Google's chief scientist, Jeff Dean, directly participated and photoshopped himself into a football player card character.

Demis Hassabis, the Nobel laureate and the founder and CEO of DeepMind, also got a personal image photo.

Netizens also showed their creativity and shared many interesting results.

Rankings

Shortly after the official launch of gemini-2.5-flash-image-preview, various lists began to show the performance of this model.

On the Artificial Analysis image editing ranking list, this model directly jumped to the first place, obtaining an ELO score of 1212.

On its text - to - image list, ByteDance's Jimeng 3.0 and OpenAI's GPT - 4o still have a slight advantage.

However, on the list of LM Arena with more votes, gemini-2.5-flash-image-preview has become the champion in both of these tasks.

The following shows more detailed scores on each indicator. Among them, gemini-2.5-flash-image-preview has obvious advantages in character consistency, creativity, charts, things/environments, etc., while in terms of stylization, GPT - 4o is currently leading.

Have you tried nano banana/gemini-2.5-flash-image-preview? How do you feel about it?

Reference links

https://x.com/googleaistudio/status/1960344388560904213

https://blog.google/products/gemini/updated-image-editing-model/

https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/

This article is from the WeChat official account "MachineHeart" (ID: almosthuman2014), author: Panda. It is published by 36Kr with authorization.