Google launches Gemini Nano Banana 2.1, its efficient AI image generation and editing model

Google has published a new model card in its Gemini API documentation for Gemini Nano Banana 2.1 (model identifier: gemini-nano-banana-2.1), described as the latest "high-efficiency" image generation and conversational editing model in the Gemini family. It is an update to Nano Banana 2, which in turn corresponds to Gemini 3.1 Flash Image, and it keeps the speed and low cost typical of the "Flash" range while introducing improvements in visual quality, prompt adherence, character consistency across multi-turn conversations and text rendering. According to the documentation itself, this model acts as the "more efficient" counterpart to Gemini 3 Pro Image, known as Nano Banana Pro, positioning the two as the Pro and Flash variants of the same Google image generation line.

Among the new features Google details for this 2.1 version, the first is improved visual quality and realism across the three available output resolutions: 1K, 2K and 4K, with 1K as the default. Tiling artifacts have also been fixed that appeared in panoramic or very elongated image formats —aspect ratios 1:4, 4:1, 1:8 and 8:1— when generated at 2K or 4K, a problem that therefore specifically affected very wide or very tall images at the higher resolutions.

The update also includes more precise text rendering and better accuracy in infographic layout, a relevant aspect for those who use this type of model to create posters, explanatory graphics or compositions with embedded text, a point that has traditionally been weak in AI-based image generators. Another improvement highlighted is multi-image fusion, which now supports up to 14 reference images in a single operation, making it possible to maintain the consistency of up to 4 different characters and the fidelity of up to 10 different objects across generations.

The model also incorporates grounding (anchoring responses in external sources) with Google Web Search and Google Image Search, which suggests it can draw on Google search results to support or verify the content it generates. It also offers configurable "Thinking" (internal reasoning) levels, with three options: minimum, medium (the default value) and high, allowing developers to adjust how much the model "reasons" before producing an image according to their speed or quality needs.

As for the technical specifications listed in the model card, Gemini Nano Banana 2.1 accepts text, image, video and PDF documents as inputs, and generates both images and text as output. Its input token limit is 131,072 and its output limit is 32,768 tokens. The documentation also specifies which capabilities it does not support: no audio generation, no caching, no code execution, no file search, no function calling, no compatibility with Google Maps grounding, no Live API support, no structured outputs and no URL context support. It does support image generation, search grounding and the Thinking feature. Regarding consumption options, the model is compatible with the Batch API, but not with the Flex or Priority inference modes.

The documentation page indicates that the stable version corresponds to the identifier gemini-nano-banana-2.1 and that this card was last updated in October 2026. Google offers the option to try the model directly in Google AI Studio and points to the general "Image generation" page of its documentation for full coverage of the features and capabilities related to image generation within the Gemini API.

Overall, this update fits into Google's strategy of maintaining two tiers of image models within Gemini —one geared toward maximum quality (Nano Banana Pro / Gemini 3 Pro Image) and another geared toward efficiency and speed (Nano Banana 2.1 / Gemini 3.1 Flash Image, in its updated version)— so that developers can choose the balance between cost, speed and quality according to the use case, whether conversational image editing, generating infographics with text, panoramic compositions or workflows that combine several reference images to keep characters and objects consistent across generations.