Google's image generation model Nano Banana 2.1 officially launched on October 6. Generating one 4K image through the API costs $0.113, and you can feed in up to 14 reference images to keep drawing the same character across scenes.
But there are complaints that quality collapses the more you edit. Let's check the pros and cons together.
What Changed in Nano Banana 2.1?
✦ Nano Banana 2.1 at a Glance
Google image generation model
Nano Banana 2.1 is Google's image generation and editing model, which reached general availability (GA) on October 6, 2026. The API model ID is gemini-nano-banana-2.1, and its predecessor is Nano Banana 2 (gemini-3.1-flash-image).
The improvements Google points to are visual design, mask-based editing, subject consistency, and more natural images. 1K is the default resolution, and it goes up to 2K and 4K.
The 4K output is said to be 4096×4096 pixels on a square basis, which is quite a resolution. It also supports extreme aspect ratios such as 8:1 wide or 1:8 tall. By the way, every output carries an invisible SynthID watermark.
You can use it in the Gemini app, Google Search AI Mode, Google AI Studio, the Gemini API, Flow, Stitch, Google Ads, and more. Not everyone gets to use it for free.
Nano Banana 2.1 has no free tier in the Gemini API, and to use 2.1 in Google Search AI Mode you must be 18 or older and use English. In Korea, Google AI Pro or Ultra subscribers are the target.
(Looks like free users can't. Even before, the Nano Banana Pro model had a lot of limits for free users.) Availability and limits in the Gemini app and AI Mode vary by product, region, and subscription tier.
How Much for One 4K Image?
✦ Price per Image
Official API pricing
By Google's official price list, image output is $30 per million tokens. One 4K image is 3,780 tokens, so $0.113 (exactly $0.1134), 2K is $0.0504, and 1K is $0.0336.
✦ What 4K Costs in Bulk
Compared with the predecessor
Make 100 images in 4K and it's about $11.34, and 1,000 images is about $113.4. The predecessor's 4K was $0.151, so on a 4K basis it got about 25% cheaper. The higher-tier Nano Banana Pro costs $0.24 for 4K, so 2.1 is less than half.
There are a few things to know. The input token price tripled from $0.5 to $1.5 per million tokens, so requests with many reference images can cost more than you'd expect.
With the default settings, thinking is on at a medium level, so the actual cost may be a bit higher than the table price, they say. Running many images together in Batch halves the price, but results take up to 24 hours. (Ugh...)
Same Character, Different Scenes
✦ 14 Reference Images
Material for keeping one character across scenes
The most useful feature this time is the 14 reference images. Put in up to 4 characters and up to 10 objects, and it tries to stay consistent even when the scene changes.
Google said the character is kept even when you edit continuously within the same conversation.
You can use it to pull a different thumbnail for each video with a YouTube channel mascot, to make product detail page scenes with the same model image in an online store, or to reduce the problem of a serial comic's protagonist face changing from panel to panel.
In Google's own published comparison, the preference score for 2.1 with thinking on was 1050, higher than the predecessor's 990 and Pro's 935. These are the company's own figures, so this is not a comparison against other companies' models.
Mask-based editing means specifying only the part you want to change, fixing it, and leaving the rest as is. (You've tried it in GPT Image 2.5, so it should feel familiar, right?)
However, there is a report that even Google admitted that errors can occur in mask editing and small text can blur.
It Draws Text Well, Too
Google said accuracy has improved for text in posters and diagrams and for infographics. In hands-on reviews too, there were comments that Chinese characters have become usable.
The first job that comes to mind is swapping the text in overseas product images into several languages.
Does It Fall Apart the More You Edit?
✦ Editing Degradation
How it falls apart
One outlet's roundup of user reviews named continuous editing degradation as the biggest drawback.
The first edit is clean, but as the count grows, noise creeps into untouched areas, colors drift, and textures warp, and several community users are said to have reported the same problem.
There is also public research showing quality dropping in repeated edits. The Banana100 paper from April made 28,000 images through 100 steps of repeated editing; the model tested was Nano Banana Pro, and noise visibly built up from the 5th to 10th edit.
It wasn't a direct test of 2.1, so it's hard to say it's the same degree, but it is a trait of the Nano Banana family.
As a case of running 2.1 directly, there is an experiment posted on October 8 to the GitHub edit-drift repository.
It's the result of one developer running it once: when editing continuously for 40 steps, the image moved further from the original at each step, and even re-running it with the conversational multi-turn method, there was hardly any improvement.
As for the cause, a guess is going around that the SynthID watermark accumulates, but the same paper found noise in open-source models such as FLUX.2 and Qwen Image Edit as well, so it saw this as not just a watermark problem.
Anyway, it seems safer to batch your requests together and, once an original you like comes out, generate anew based on it. Dragging out an image editing conversation for long with this model is not a good choice.
How Does It Stack Up Against GPT Image 2.5?
✦ Arena Text-to-Image Ranking
As of October 7, early tally
On Arena's official text-to-image leaderboard (as of October 7), 2.1 is 6th with 1,328 points. 1st and 2nd are the two versions of GPT Image 2.5 (1,425 and 1,398 points) and 3rd is GPT Image 2 (1,383 points), so it's nearly 100 points behind 1st.
Still, it's the highest among Google's image models. The predecessor Nano Banana 2 is 11th and Pro is 16th. It's an early tally right after launch, so the ranking can change.
From the reviews, there's also an assessment that GPT Image 2.5 is still ahead in overall quality. So it's right to split it up: the finish of a single image goes to GPT, character consistency and price go to 2.1.
Who Will Find It Useful?
If you're a blogger or YouTuber who makes several images a day, you have plenty of reason to switch from the tool you're using now. The price per image is low and the character holds, so it's easy to match your channel's look.
On the other hand, if you're polishing a single illustration with care, because of the editing degradation you need the habit of finishing after the first few fixes.
If you need to churn out ad creatives in bulk, start with the cost calculation. Make 300 4K images a day for 30 days and it comes to over $1,000 a month. That's not a small cost.
Developers, Look After the Predecessor First
✦ Predecessor Timeline
After the 2.1 launch
With 2.1, Google designated it as the recommended replacement for the predecessor Nano Banana 2 (gemini-3.1-flash-image).
A report went around that it would close on October 29, but Google's official deprecation schedule has not yet announced an end date for the predecessor.
Anyway, before a date is confirmed and suddenly announced, it would be good to try the new model ahead of time and get the migration work done.
To sum up, Nano Banana 2.1 is not a model that overwhelmingly beats the others. It's a model that's useful in consistency, text, and price.
If you decide where to use it before going in, satisfaction is high, and if you hand it work that needs many revisions wholesale, you're likely to be disappointed because of degradation. Anyway, as consumers like us, it's good to be given so many options.
I feel GPT Image models are still the best, but I hope better rival models come out.
That's all for this post. See you next time.
🌐 Korean version: 4K 이미지 한 장에 약 150원.. 구글 나노바나나 2.1 출시! 쓸만할까?
No comments:
No comments yet. Want to leave the first one?
Post a Comment