Google DeepMind just dropped a new family of image generation models with a name that sounds like it was coined during a late-night brainstorming session fueled by actual bananas. The Nano Banana 2 Lite, officially branded as Gemini 3.1 Flash Lite Image, has landed at the fifth spot on text-to-image leaderboards with an Elo score of 1,255 in evaluations from Artificial Analysis.

It also placed ninth in Multi-Image Edit rankings. For a model designed to be the budget-friendly option in the lineup, that’s a surprisingly strong showing.

What the Nano Banana 2 series actually does

The Nano Banana 2 family launched around February 26, 2026, and it comes in multiple variants. The Lite version is positioned as the fastest and most affordable option, aimed squarely at developers and businesses running high-volume image generation tasks.

The feature set across the series is genuinely comprehensive. Conversational multi-turn editing lets users refine images through back-and-forth dialogue rather than starting from scratch each time. Variable aspect ratios mean you’re not locked into square outputs. And upscaling goes all the way to 4K resolution, which puts it in the range of production-quality visual content.

Text rendering, historically one of the weakest points for AI image generators, is listed among the supported capabilities. Anyone who has watched an AI model butcher the word “restaurant” on a storefront sign knows why this matters.

Every generated image gets an invisible SynthID watermark baked in. This is Google’s approach to the growing concern around AI-generated content being passed off as authentic photography or artwork.

On the pricing front, the main variant uses a token-based system. Generating a standard 1K output image requires approximately 1,120 tokens.

The competitive landscape in AI image generation

The Pro version of Nano Banana 2 has also been appearing in the top five to seven positions on these same benchmarks, suggesting Google has managed to build a lineup where even the economy option punches above its weight.

The series also integrates real-world knowledge through web search functionalities, allowing the model to pull in contextual information from the web to improve how accurately it represents real-world subjects, landmarks, or concepts.

Subject consistency and instruction adherence were explicitly targeted for improvement in this generation.

And before anyone gets confused: no, this has nothing to do with the meme token called Nano-Banana (NANOBANANA) on the Solana blockchain. The naming overlap is purely coincidental.