জেমিনি 3.1 ফ্লেছ লাইট ইমেজ

The lite tier of Gemini 3.1 Flash Image, generating and editing images at half the image-output price of the Flash model it is based on. It keeps the same aspect-ratio and resolution controls, trading some fidelity for cost on high-volume work.

Nano Banana 2 Lite is the cheapest and fastest member of Google's image family, and the lab is unusually direct about the job it is for: rapid ideation and high-velocity pipelines where speed and cost are the constraints that decide the design. Google puts a number on the speed - a text-to-image result in about four seconds - which is the difference between a person waiting for a picture and a person flipping through options.

The two figures that define it are latency and price. Google quotes roughly three and a half US cents for a 1K-resolution image, which is what makes it viable to generate thousands of drafts, or to put generation in front of users on a budget that does not scale with enthusiasm. The lab's own recommendation is blunt: developers still on the original Nano Banana should swap to this model, because it is better, faster and cheaper on every dimension Google measures.

What survives the compression matters more than what was cut. Google says it keeps reliable prompt adherence, strong character consistency and legible in-image text - the three things that usually break first when an image model is optimised for speed, and the three that decide whether a draft is worth iterating on at all.

The ladder Google published with this release is the clearest statement of where each model belongs. This one is built for speed, for near-real-time and high-volume work where latency is critical. Nano Banana 2 is the generalist that balances quality against cost. Nano Banana Pro handles complex professional work where accuracy matters more than speed. The original Nano Banana is named as legacy.

The trade is fidelity and headroom. It generates a single 1K image per call, without the 2K and 4K output of its larger siblings, and Google routes accuracy-critical work to the Pro model rather than this one. Editing is supported and directed in language: there is no mask, so a change is described rather than painted onto a region.