Google DeepMind Claims Nano Banana 2.1 Beats Gemini 3 Pro Image on Preference Benchmarks
Summary
Google DeepMind publishes the model card for Nano Banana 2.1, an image generation and editing model in the Gemini 3 series built on Gemini 3.6 Flash. The company claims it beats Gemini 3 Pro Image on preference benchmarks, reporting 1050 versus 935 Elo overall, while citing weaknesses in small text, character consistency, and spatial localisation.
Key Points
- Google DeepMind publishes the model card for Nano Banana 2.1 on 6 October 2026, an image generation and editing model in the Gemini 3 series built on Gemini 3.6 Flash.
- The model accepts text and images with a 1M token context window and outputs 4K token images and 64K token text through the Gemini App, AI Studio, Gemini API, Search AI Mode, Ads, Flow, and Stitch.
- DeepMind reports Nano Banana 2.1 outperforms Gemini 3 Pro Image on benchmarks like Overall Preference at 1050 versus 935 Elo, while noting limitations in small text rendering, character consistency, and spatial localisation.