Nano Banana 2.1 tested: half the price, and no made-up text
We made 32 images with Nano Banana 2.1 and Nano Banana 2. Both wrote all asked-for text right, but only Nano Banana 2 added made-up claims like "100% pure". And 2.1 costs half.
Google has a new image model: Nano Banana 2.1. It makes and edits pictures, and Google says it writes text in images better than before. For developers it is also half the price of Nano Banana 2. We made 32 images with both models to see what changed. Both wrote every word we asked for correctly. But Nano Banana 2 often added text we never asked for, like "100% pure & natural" on a honey jar. Nano Banana 2.1 did that in none of its images.
What is Nano Banana 2.1?
Nano Banana is the nickname of Google's image model in the Gemini app. You type what you want, and it makes a picture, or changes a photo you give it. Version 2 came out in February 2026. Version 2.1 is new: Google updated its developer pages for it on 6 October 2026.
Google calls 2.1 its "most advanced image model", with better quality, reasoning and design, on its Gemini page.
What's new, according to Google
Google's model page lists what changed compared to Nano Banana 2:
- More realistic images, at all three sizes (1K, 2K and 4K).
- No more repeated "tiles" in very wide or very tall images, like panoramas.
- Better text in images, and better layouts for infographics. An infographic is a picture that explains something with icons and short text.
- Up to 14 example images to work from, so the same people or objects can return in a new picture.
- A setting for how long the model thinks before it draws.
Google also says the same character stays the same better when you ask for changes one step at a time.
Who can use it?
In the Gemini app, Nano Banana 2.1 needs a paid Google AI plan. Free users still get Nano Banana 2, Google's Gemini page says. Paid plans also get more images per day, but Google does not say how many.
Developers can use it through the Gemini API, the connection that programs use, and through services like OpenRouter. That is how we tested it. In the Gemini app, every image also gets a hidden SynthID watermark and a visible mark, so people can see it was made by AI, Google says.
What does it cost?
For developers, Google's price list gives these prices per image:
- Nano Banana 2.1: $0.0336 for a standard 1K image, $0.0504 for 2K and $0.113 for 4K.
- Nano Banana 2: $0.067 for 1K, $0.101 for 2K and $0.151 for 4K.
So for 1K and 2K images, the new model costs half as much. For 4K the difference is smaller, because a 4K image from 2.1 uses more tokens. Tokens are the small pieces that AI models count to work out the price. In the Gemini app you do not pay per image; you pay for the plan.
Our test: text in images
Text in pictures used to be a weak point for AI. So we gave both models four prompts with exact text, and made each one twice:
- A bakery sign: "Bakkerij De Molen" and "Open 7:00 – 18:00".
- A chalkboard menu with five items and prices, like "Appeltaart €4,10".
- An infographic titled "Save Water at Home" with three numbered steps.
- A honey jar with "Wild Flower Honey", "350 g" and "Best before 12/2027".
That is 14 texts per try. We read every image ourselves. Both models got all 28 texts right. Capital letters we did not count as mistakes.
The difference was in what they added. Nano Banana 2.1 wrote only what we asked for, in all 8 images. Nano Banana 2 added extra text in 6 of its 8 images.
Nano Banana 2 adds text you did not ask for
Some of that extra text is harmless, like "BROOD & GEBAK" (bread and pastry) above a shop window. But some of it sounds like a fact. On the bakery's window, Nano Banana 2 wrote "SINDS 1923" (since 1923). On the honey jars it wrote "Raw & Unfiltered" and "100% PURE & NATURAL".
For a fun picture that does not matter. For a real product photo or a real shop, it does. A label that says "100% pure" is a claim, and you are the one who has to be able to prove it. So check every word in an AI image before you use it, especially with Nano Banana 2.
Infographics: short and clean, or full of extra tips
For the water infographic, we asked for a title and three short steps. Nano Banana 2.1 made exactly that, twice: a clean picture with three icons and the three steps.
Nano Banana 2 made busier infographics. It added a line of advice under every step, like "Aim for 5-minute showers". In one try it also added "Reduces water use by up to 50% compared to a bath" and "Leaks waste thousands of gallons yearly". We did not ask for those numbers, and we do not know where they come from. If you share an infographic like that, those claims become yours.
The same person, three scenes
Next we asked for a woman with short red hair and freckles in a yellow raincoat, in the rain. Then we asked for the same woman in a café, and then on a bike along a canal. Each step started from the picture before it.
Nano Banana 2.1 kept her look the same in both tries: same short red hair, same freckles. Nano Banana 2 did well in the first try. In the second try, the first picture showed a woman with curly, darker hair, and the next two pictures had short straight red hair. So it changed the person halfway.
Panoramas: no difference we could see
Google says 2.1 fixes repeated "tiles" in very wide images. We asked both models for a panorama of a Dutch canal street at sunset, eight times wider than high, at 2K.
In our four panoramas we saw no repeated blocks in either model. Both looked good. Maybe the fix shows more at 4K or with other scenes. With our test, we cannot show a difference here.
Speed and cost in our test
Through OpenRouter, a standard image from Nano Banana 2.1 cost us $0.036 on average, and one from Nano Banana 2 $0.067. That matches Google's list prices. The wide 2K panoramas cost $0.053 and $0.101.
Nano Banana 2.1 was also a bit faster: in the middle 8.4 seconds per image, against 9.3 seconds for Nano Banana 2. For the panoramas it was 13 against 18 seconds. All 32 images together cost $1.75.
How we checked
On 9 October 2026 we read Google's model page, its price list and the Gemini image page. The screenshots are our own captures of those pages, cropped only.
We made the images the same day through OpenRouter, with Google AI Studio as the provider: 8 prompts, 2 models, 2 tries each, all with default settings. We read every text and compared the faces by eye, so part of the result is our judgement. The test images in this article were made by AI, as part of the test; the other charts were made by us. The owner of this site paid for the images.
This was a small test with clear prompts. With other prompts, the results may differ. Earlier we also compared AI video models with the same prompt.