What Nano Banana Actually Is, and Which One You Mean
Ask "what is Nano Banana" and you get four answers, because Google now ships four models under that name. Only one of them is the original. That one dies on 2 October 2026.
We run all four on our API, so we field this question constantly. Here is the short version: Nano Banana is Google's gemini-2.5-flash-image, a fast 1K image model that earned its reputation in 2025 and is now labelled legacy by the company that made it. Everything else wearing the name is newer, cheaper, or both.
One name, four different models
The naming is genuinely confusing, and not by accident — "Nano Banana" became a brand before it became a product line. What it covers today:
| Name | Google model ID | What it is |
|---|---|---|
| Nano Banana | gemini-2.5-flash-image | The original. Retires 2 October 2026 |
| Nano Banana 2 | gemini-3.1-flash-image | Current generation, up to 4K |
| Nano Banana 2 Lite | gemini-3.1-flash-lite-image | Fastest and cheapest, 1K only |
| Nano Banana Pro | gemini-3-pro-image | Premium tier, role-separated references |
Two traps in that table. Nano Banana 2 and Nano Banana Pro are not the same model — some catalogs list Pro as "Nano Banana 2" and that is simply wrong; they are built on different weights and priced differently. And Lite is newer than the unsuffixed original, which reads backwards but is true: gemini-3.1 versus gemini-2.5.
If you want the full "which one for which job" breakdown, we wrote that separately: Nano Banana vs Nano Banana Pro, 2 and Lite. This page is about the original.
What the original model actually is
gemini-2.5-flash-image is a flash-tier model, which in Google's vocabulary means optimised for speed and cost rather than maximum quality. It generates from a text prompt and it edits from reference images. Both capabilities live behind the same weights.
The concrete specifications:
- Resolution: 1K, and only 1K. There is no 2K path and no 4K path on this generation. If a client needs print output, this was never the model.
- Ten aspect ratios — 21:9, 16:9, 4:3, 3:2, 1:1, 9:16, 3:4, 2:3, 5:4, 4:5. Enough for a banner set and a story crop from one prompt template.
- Up to three reference images when editing. Google's own guidance: include a maximum of three images in an input for best results.
- 65,536 tokens in, 32,768 out, with a knowledge cutoff of June 2025.
- SynthID watermark on every output, always, with no opt-out at any provider.
Speed was its selling point and still is. OpenRouter's public numbers put its median at roughly 1.94 seconds through AI Studio and 4.48 through Vertex. Our own pipeline budgets roughly 30 seconds end to end, because that number includes queueing and delivery on our side rather than raw inference.
We price it at $0.0312 per image, against $0.039 at Google's own API, fal.ai and OpenRouter as of our 26 August 2026 check.
Prices and product terms can change. Check each provider's current pricing before making a purchasing decision. Google Batch or Flex pricing is not directly equivalent to a standard on-demand API request because scheduling, availability, and processing conditions differ; it is therefore excluded from this comparison. This is a scoped comparison, not a claim that E2X is the world's cheapest option in every configuration.
The three things it cannot do
Every model page that sold you this one glossed over these. We would rather you hear them now.
It cannot write legible text inside an image. That capability landed with Gemini 3; this is a 2.5-generation model. Ask for a poster headline and you get letterforms that survive at thumbnail size and fall apart at full size. Packaging mockups, infographics, menu boards — wrong model, every time.
It cannot go past 1K. Not with a parameter, not with a flag. The ceiling is the ceiling.
It cannot hold a complex reference composite. Three images is the working limit. The newer flash models take fourteen.
None of these are bugs. They are what a 2025 flash-tier model was built to be, and the model was excellent at the job it was scoped for.
Google's own verdict
The vendor is not being subtle about where this goes. Straight from the image generation docs:
"The legacy pioneer of the Nano Banana series. While it has been a reliable workhorse, we strongly recommend that customers transition to Nano Banana 2 Lite to experience enhanced quality, faster generation speeds, and lower API pricing."
Notice the source. That is the vendor describing its own product, published on the same docs site as the table that gives the model an end date.
"Strongly recommend" plus a hard date is what an eviction notice looks like in model-card language.
Calling it while it is still here
The contract is the same as every model in our catalog: submit a job, then poll it or take a webhook.
curl -X POST https://api.e2x.ai/v1/jobs/submit \
-H "Authorization: Bearer $E2X_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "google/nano-banana/text-to-image",
"input": {
"prompt": "A weathered enamel coffee pot on a camp stove at first light, steam catching the sun",
"aspect_ratio": "3:2"
}
}'
One schema detail catches almost everyone: our aspect_ratio default on this model is 9:16, not 1:1. Portrait unless you say otherwise. Set it explicitly and stop guessing.
Editing works through the same model's edit endpoint, which takes image_urls alongside the prompt. Remember the three-image ceiling.
What to do before October 2
Treat the date the way you would treat any dated dependency: a scheduled incident with a known fix. When the weights go dark, every request to this slug fails at the same moment, everywhere at once. Nobody can route around a vendor retirement.
The replacement is Nano Banana 2 Lite, and the migration is a one-line change:
- "model": "google/nano-banana/text-to-image",
+ "model": "google/nano-banana-2-lite/text-to-image",
We charge $0.0238 for Lite against $0.0312 for the model it replaces. Cheaper and a full generation newer — which almost never happens, and removes the only argument anyone had left for staying put.
Two things to check while you are in there. Lite's aspect_ratio default is 1:1, not 9:16, so pin it explicitly or every asset quietly becomes a square. And Lite has no resolution parameter at all — drop the field if you were sending one.
If you need more than 1K, go to Nano Banana 2 instead. If your images carry readable text, Nano Banana Pro is the tier where Google put that capability. The rest of what we run is in the text-to-image category and the full catalog.
Frequently asked questions
What is Nano Banana?
Nano Banana is the product name for Google's gemini-2.5-flash-image, a fast image generation and editing model released in 2025. It outputs at 1K resolution, supports ten aspect ratios, and accepts up to three reference images when editing. Google now labels it legacy and retires it on 2 October 2026.
Is Nano Banana the same as Nano Banana 2?
No. Nano Banana is gemini-2.5-flash-image; Nano Banana 2 is gemini-3.1-flash-image, a full generation newer with 4K output, fourteen reference images and legible in-image text. They are separate models with separate prices. Some catalogs also mislabel Nano Banana Pro as "Nano Banana 2" — that is a third, different model again.
When does Nano Banana shut down?
The date sits in Google's own deprecation table, against gemini-2.5-flash-image: 2 October 2026. Once that day passes the endpoint stops answering — here and at every other provider, since the weights are Google's to switch off. Move before it, not on it.
Can Nano Banana render text inside an image?
Not reliably. Legible in-image text arrived with Gemini 3 and this model is a generation older; a single word might come out clean, a sentence rarely does. Anything with a headline in it belongs on Nano Banana Pro.
How much does Nano Banana cost on E2X?
We bill $0.0312 an image. Google's own API, fal.ai and OpenRouter all sat at $0.039 on 26 August 2026, so that puts us 20% under the going rate. The replacement, Nano Banana 2 Lite, runs $0.0238 here.
Do Nano Banana images have a watermark?
Yes. Every image comes out carrying SynthID, the invisible provenance mark Google applies to its AI output. There is no flag for it anywhere, and a provider advertising one would be selling something it cannot deliver. Settle it with legal or brand before you integrate, not after.
What should I migrate to?
Nano Banana 2 Lite, in almost every case. It is cheaper than the original, a generation newer, and Google names it as the recommended replacement. Change the model slug — the request and response shapes are unchanged. Go to Nano Banana 2 instead if you need output above 1K.