Nano Banana 2
Nano Banana 2 is gemini-3.1-flash-image: up to 4K, ten object reference images, and $0.067 for a 1K image. It sits at rank 7 on the Arena text-to-image board while Nano Banana Pro sits at rank 12, so the cheaper Google model is the higher ranked one.
- up to 4K
- ten object references
- $0.067 per 1K image
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
A real gap once, a tie the other time
On the text-to-image board Nano Banana 2 scores 1263.8 and Nano Banana Pro scores 1246.0. The gap is about eighteen points and the two confidence intervals do not touch: this model starts at 1258.7 and Pro ends at 1249.4. So that is a real difference, not noise.
On the editing board the story runs the other way and should be written honestly: Pro has 1389.1 and this model has 1385.1, so Pro is four points ahead. But the intervals overlap completely (1381.0 to 1389.1 against 1386.1 to 1392.0), so on editing these two are level and nobody can say which is better.
Now put the price beside those two results. A 1K image is $0.067 here and $0.134 on Pro. Exactly half, for output that wins once and ties once.
The catch: half the price is only true at one size
This model has four size tiers and Pro has two. Line them up: at 0.5K this model is $0.045 and Pro has no such tier at all. At 1K it is $0.067 against $0.134, so half. At 2K it is $0.101 against the same $0.134, because Pro charges one price for both 1K and 2K. At 4K it is $0.151 against $0.24.
So the bigger you generate, the smaller the saving gets: half the price at 1K, about three quarters at 2K, about sixty percent at 4K. Our price column takes the 1K tier because it is the only tier both models publish a number for. If your work is always 2K, the saving you see in our table is larger than the saving you will see on your invoice.
Ten reference images, and the two numbers standing next to it
Google publishes three separate ceilings for this model: up to ten object images, up to four character images, and up to three style images. Our reference column takes the object number, and on that number this model tops the table. But if your work is several consistent people in one scene, the number that matters to you is four, and Pro takes five. That is the only published, citable advantage Pro holds.
Which configuration actually earned the rank
The Arena row is named "gemini-3.1-flash-image (nano-banana-2) [web-search]". That suffix is not decoration: the number we put in our table belongs to the search-grounded configuration. Call the same model without grounding and you have not bought exactly what we measured. This is the pattern we keep running into across this reference and keep writing on every model page: a leaderboard row names a serving configuration, not only a model.
The Lite version people confuse with this one
Google has a third model, gemini-3.1-flash-lite-image, which appears on Arena as Nano Banana 2 Lite at rank 11. It only does 1K, takes no character or style references at all, and Google itself writes that it is not optimised for multiple reference inputs or multi-turn sequential editing. In exchange it accepts up to fourteen object images, the highest of all three. We did not make it a candidate. If you want the cheapest possible route, go and look at it knowing what you give up.
From Iran
We read the Gemini API available-regions page today and Iran is not on it. That statement is about the API and says nothing about the Gemini app. We do not sell accounts and we do not suggest ways around it.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- blocked
- Payment
- no working route
- Free tier
- no
- Measured on
How we checked: The Gemini API available-regions page lists the countries where the API works, and Iran is not on that list. That is reading a page, not measuring a network.
What it is good at
- Ranked above Google own flagship on text-to-image, by a margin the confidence intervals support
- Half the price of Pro at the 1K tier, and cheaper at all three other tiers
- Ten object reference images, the highest of any candidate in the table
- It has a 0.5K tier that Pro does not, for when a small output is enough
- Google describes its text rendering as advanced, for infographics, menus and diagrams
Where it falls short
- It is not ahead of Pro on editing: four points behind, with intervals that overlap completely, which means a tie.
- Half the price is only true at 1K; by 4K the gap narrows to about sixty percent.
- It takes four character references where Pro takes five.
- Google publishes no pixel dimensions for any tier from 0.5K to 4K.
- No weights are published and there is no local route.
- Iran is not on the Gemini API available-regions list.
- Our number comes from the web-search row, so it is tied to a configuration and not only to the model.
Our take
Between the two Google image models this is the default: cheaper, ahead on text-to-image, level on editing. Reach for Pro when you genuinely need that fifth character reference, which is a specific job most people do not have. And if editing is the real work, the top row of this table is not a Google model at all.
Questions people actually ask
What is Nano Banana 2
It is the marketing name for Google image model gemini-3.1-flash-image. It generates up to 4K, accepts up to ten object reference images, and ships on the ordinary Gemini API.
How much does Nano Banana 2 cost
On the standard tier: $0.045 for 0.5K, $0.067 for 1K, $0.101 for 2K and $0.151 for 4K. The batch route is roughly half of those numbers.
Is Nano Banana 2 better than Nano Banana Pro
On text-to-image yes, and the gap sits outside the confidence intervals. On editing the two are level. And at 1K it costs half what Pro costs.
What is the difference between Nano Banana 2 and the Lite version
Lite only does 1K and takes no character or style references, but accepts up to fourteen object images. Google itself writes that it is not optimised for multi-turn editing.
Is it available from Iran
Iran is not on the Gemini API available-regions list, so there is no official route. We do not sell accounts and we do not suggest ways around it.
Sources
- Google, image generation in the Gemini API vendor source 12 August 2026
- Google, Gemini API pricing vendor source 12 August 2026
- Google, Gemini API available regions vendor source 12 August 2026
- Arena, text to image leaderboard vendor source 12 August 2026