In August 2025, a mysterious model named "nano-banana" appeared on LMArena. It was later revealed to be "Gemini 2.5 Flash Image" released by Google. Rather than being a diffusion-based model, it utilizes an autoregressive method to generate 1,290 tokens.
Model ReleasesGoogleGemini 2.5 Flash ImageNano Banana
Google Nano Banana Changes AI Image Generation Standards with Strict Prompt Adherence and Low Cost
This article is a translation. Read the Japanese original
The author of the original article prioritizes fidelity to specific prompts over aesthetic beauty in rankings. Nano Banana is evaluated as accurately following extremely complex instructions thanks to its powerful text encoder. The author suggests that Google may be underestimating the model's performance in terms of this precision.
Available methods for use include the Gemini app, the "Create Image" tool on the web version, and Google AI Studio. These are available for free, although a watermark is placed in the bottom right of the generated images.
For developers, the "gemini-2.5-flash-image" endpoint is provided via the Gemini API. The cost to generate a 1-megapixel image is approximately $0.04. This is comparable to many diffusion models and cheaper than the $0.17 for gpt-image-1. Additionally, "gemimg," a Python package that simplifies the tedious processes involved in API usage, has been released as open source.
Source: Nano Banana can be prompt engineered for nuanced AI image generation（HN 887pt・236コメント） (HN Search (backfill), 2025-11-14)