What Is Nano Banana? Google's Image Editing Model and Its Four Variants
Nano Banana is the nickname that stuck to Google's image editing models. Which variant is which, what they cost per image, and why it is the cheapest step in most video pipelines.
"Nano Banana" started as a nickname and became the name people search for. It refers to Google's image generation and editing models, and by 2026 it covers several distinct variants that get conflated constantly. Here is the map.
The variants
| Name | Also known as | Endpoint | Price |
|---|---|---|---|
| Nano Banana (original) | Gemini 2.5 Flash Image | fal-ai/gemini-25-flash-image/edit | — |
| Nano Banana 2 | — | fal-ai/nano-banana-2/edit | $0.08/image |
| Nano Banana Pro | Gemini 3 Pro Image | fal-ai/nano-banana-pro/edit | $0.15/image |
| Nano Banana Lite | — | google/nano-banana-lite/edit | — |
Prices read live from fal.ai on 2 August 2026. The Pro variant is the same model as Gemini 3 Pro Image — two names, one thing, which is where most of the confusion comes from. Lite is the latency-optimised member of the family, sub-2-second generation with 14 supported aspect ratios.
What the schema tells you
Nano Banana Pro's edit endpoint takes image_urls (one or more inputs) plus a prompt, and offers some parameters worth knowing:
resolution— 1K, 2K or 4K, defaulting to 1K. If you are feeding the output into a video model, 1K is usually enough; video models downsample anyway.aspect_ratio— defaults toauto, inheriting from the input.system_prompt— an optional instruction steering persona and output style across the whole request. Rarely used, genuinely useful for keeping a series consistent.enable_web_search— lets the model pull current information into the generation. Unusual for an image model and occasionally exactly what you need.safety_tolerance— 1 (strictest) to 6, defaulting to 4.seed— for reproducibility, which matters if you are iterating on one composition.output_format— png by default; jpeg and webp available.
Why it matters for video work
Because it is the cheap half of most good pipelines.
Video generation costs $0.10–0.26 per second. Image editing costs 8 to 15 cents per attempt. Any decision you can move from the video step to the image step gets roughly twenty times cheaper to iterate on.
Concretely: for the AI hug format, compose the two people into one frame with Nano Banana first, approve the still, then animate it. For a character series, design the character as an image and reuse it. For a product effect, get the product shot right before adding motion.
The general rule: decide composition in images, decide motion in video. People who skip this iterate at $1.40 a try instead of $0.08.
Which variant to use
Nano Banana 2 ($0.08) for volume iteration — trying twenty compositions, designing a character, ordinary edits.
Nano Banana Pro ($0.15) when the edit is hard: reconciling two photographs with different lighting, complex multi-subject composition, anything where the first model keeps producing anatomically confused results. Roughly double the price and worth it on the difficult cases.
Lite when latency matters more than fidelity — interactive or high-throughput contexts.
The limits
It is an image model. It does not animate, and a still is not a video no matter how good it is. And the same likeness considerations apply as everywhere else: putting a real person into a composed scene is a decision about that person, not just about the prompt.