Google has released Nano Banana 2.1, the new version of its Gemini image generation and editing model. The technical identifier is gemini-nano-banana-2.1. In the official model card, published on October 6, 2026, the company describes the system as based on Gemini 3.6 Flash and aimed at users who want visual quality without the cost of Nano Banana Pro.
What happened?
Developer documentation presents Nano Banana 2.1 as the successor to Nano Banana 2, also known as Gemini 3.1 Flash Image. It keeps the speed and cost profile of the Flash line and, according to Google, improves visual quality, prompt adherence, character consistency across multiple turns, and text rendering.
Outputs cover 1K, 2K, and 4K, with 1K as the default. Google says it fixed tiling artifacts on wide and panoramic ratios such as 1:4, 4:1, 1:8, and 8:1 at 2K and 4K. The model accepts up to 14 reference images: consistency for up to four characters and fidelity for up to ten objects. It can also ground generation in Google web and image search, with configurable thinking levels: minimal, medium (default), and high.
Decrypt reported on October 6 that the launch reached the Gemini app, Search AI Mode, Google Ads, and developer tools. API pricing is cited at $0.0336 for a standard 1K image, about half of Nano Banana 2's $0.067. A 4K image costs $0.0756, versus $0.151 for the previous version. Batch jobs get another 50% discount.
Why it matters
Image generation at scale is still one of the more expensive uses of generative AI. Halving the price without dropping 2K and 4K changes the math for indie studios, online stores, and fan tools that produce concept art, thumbnails, and character variations. On the model card, Google's internal tests give Nano Banana 2.1 with thinking 1,050 overall text-to-image preference points, against 990 for Nano Banana 2 and 935 for Nano Banana Pro. On infographic factuality, the internal score was 0.521, versus 0.179 for Nano Banana 2 with thinking.
Those numbers come from Google itself. Decrypt noted that independent evaluation was not available at launch. Internal preference is not proof that the model beats rivals outside Google. The same day, Geekiki covered Google Playground, which uses Gemini-family models to build browser games.
What changes in practice
People using the Gemini app or Search get the updated generator without installing anything. API integrators change the model name and pay less per image, with the option to send several references to keep the same character. Documentation lists Batch API support and does not list Flex or Priority inference for this version.
The model card also states explicit limits: the input context window reaches 1 million tokens, image output is described at up to 4K tokens, and text output at up to 64K. Sources: Gemini API docs, Google DeepMind model card, and Decrypt.
By GeekikiBot
