Google's Nano Banana 2.1 slices image-gen API price in half, tops ELO charts 🍌

Google on Tuesday released Nano Banana 2.1, the newest version of its AI image generation and editing model, rolling it out across the Gemini app, Google Search's AI Mode, Google Ads and developer tools including Google AI Studio, Flow and Stitch. The update delivers what Google described as notable gains in visual design, mask-based editing and subject consistency, the last of which helps people and objects retain their identity across multiple edits. The model also retains the search-grounding capability introduced in version 2, allowing it to consult Google Search and Google Image Search before rendering real-world subjects.

On internal text-to-image preference tests using the ELO scoring system, in which higher points indicate stronger human preference with no fixed ceiling, Nano Banana 2.1 scored 1,050 ELO points, ahead of Nano Banana 2 at 990 and Nano Banana Pro at 935. The model accepts up to 14 reference images, tracks up to four characters and ten objects, outputs resolutions up to 4K and supports aspect ratios as wide as 8:1. Developers can configure a "thinking" setting ranging from minimal to high before the model begins rendering.

Pricing through Google's developer API was also cut sharply. A standard 1K image now costs $0.0336, roughly half of Nano Banana 2's $0.067, while a 4K image runs $0.0756 versus $0.151. Batch processing, in which large volumes of images are generated together rather than individually, receives an additional 50% discount, putting the effective cost of a thousand standard 1K images at about $33.60 compared with $67 previously. Google framed the release as both a quality and efficiency upgrade.

The original Nano Banana, launched earlier in 2025, helped propel Gemini to the No. 1 position on both major U.S. app stores in September 2025, displacing ChatGPT after nearly three years at the top, and contributed to Alphabet's market capitalization surpassing $3 trillion during the same period. Version 2 followed in February, built on the Gemini 3.1 Flash Image architecture. The Tuesday announcement did not alter that underlying stack.

Separately, Paris-based Mistral AI launched Mistral Large 4 on Tuesday, a 1 trillion-parameter model it described as the strongest open-weight system from the United States or Europe, with Chief Scientist Guillaume Lample stating, "[Mistral Large 4] is at the frontier of open models, and by far the strongest open-weight model from the US or Europe."