Nano Banana 2 Lite Guide: Speed, 1K Output and Cost at Volume

Most image-model comparisons rank on quality. For a large class of products, that is the wrong axis. A sticker maker, a thumbnail generator or a live "restyle this photo" button lives or dies on how long the user waits, and an image that takes eight seconds feels broken no matter how good it is.

Gemini 3.1 Flash Lite Image, which Google also calls Nano Banana 2 Lite, is built for that class. It reached general availability on 30 June 2026 as gemini-3.1-flash-lite-image, and it gives up a lot to be fast and cheap. Knowing exactly what it gives up is the whole decision.

What Lite gives up

Google's model page is direct about the limits:

• 1K output only. 1024x1024, set with image_size. Google states that 2K and 4K are not supported.
• A small output budget. 4,096 output tokens per response, against a 65,536-token input limit. Some catalogues list a far larger output cap; Google's figure is the one that applies.
• No search grounding, so it cannot pull in reference images of real subjects the way Gemini 3.1 Flash Image can.
• No structured outputs, function calling or caching.

What it keeps is more than the name suggests. It accepts text, images, video and PDF as input, supports interleaved text-and-image output, handles multi-turn local edits, and offers 14 discrete aspect ratios. Thinking is available at two levels, minimal and high, so you can trade a little latency for better adherence on harder prompts. Every output carries a SynthID watermark plus C2PA provenance data.

Latency is the product

Google says the model targets sub-2-second end-to-end latency, with significantly reduced TPU compute cost. Treat that as a target, not a guarantee: network distance, prompt length, input images and thinking level all add time.

Still, the target changes what you can build. Under about two seconds, an image feature can sit inside an interactive flow — type, see, adjust, see again. Above five, it becomes a job the user submits and comes back to. If your interface is the first kind, measure time to a usable image on Lite before considering anything heavier. Latency-adjusted model scoring covers how to weigh speed against quality in one number.

Keep thinking at minimal for interactive paths. Switch to high only where you have measured a quality gain worth the delay, such as edits with several constraints.

Cost at volume

Google's pricing page lists image output at $30 per million tokens, which works out to $0.0336 per 1K image. Text and image input is $0.25 per million tokens and text or thinking output $1.50 per million. The Batch API halves everything, to $0.0168 per image.

At 100,000 images a month, that is roughly:

• Lite, standard — $3,360
• Lite, batch — $1,680
• Flash Image at 1K — $6,700
• The original Nano Banana at $0.039 — $3,900

Input costs come on top and depend on how many reference images you send. The gap to Flash Image is the number that matters: Lite is half the price at the same resolution, so Flash has to be visibly better on your prompts to earn its place. One caveat: Google's pricing page lists Nano Banana 2.1, the October update to Flash Image, at the same $0.0336 per 1K image, so against 2.1 the case for Lite rests on latency rather than price.

Replacing the original Nano Banana

Google's deprecations page names Lite as the recommended replacement for Gemini 2.5 Flash Image, and the fit is natural: both output at 1K and both were built around fast conversational edits. Google says Lite maintains high character alignment with the original Nano Banana's standards, and it is slightly cheaper per image.

The differences to check are the ones that break quietly. Lite adds a thinking mode, so confirm your response parser tolerates any extra parts. Its aspect-ratio list may not match exactly what you requested before. And safety filtering varies between generations, so rerun prompts near the edge of your content policy.

Where it fits, and where it does not

Good fits:

• Sticker and avatar generators, where the output is small anyway.
• Background swaps, colour changes and other quick local edits inside an editor.
• Thumbnails and placeholder art generated per item in a catalogue.
• Rapid exploration: generate many drafts cheaply, then re-render the chosen one on a higher-resolution model.

Poor fits:

• Anything printed or shown large. 1K is a hard ceiling, and upscaling adds its own artefacts.
• Images of specific real-world subjects that need grounding to be accurate.
• Dense text in images, infographics and diagrams, where Google markets the larger Flash model on text rendering.

The draft-then-render pattern is often the best of both. Lite produces four options for about $0.13; the chosen one goes to Flash Image at 2K or 4K. That keeps the interactive part fast and spends the premium only on the image the user actually wants.

A quick evaluation

1. Take twenty prompts from your product, including your five most common edits.
2. Run each on Lite with thinking at minimal, then at high, and on Flash Image at 1K.
3. Record wall-clock time to the final image from your own servers, not the vendor's figure.
4. Have someone rate the outputs blind. If Lite is acceptable on most prompts, make it the default and route the rest to Flash.