Google releases Nano Banana 2.1: better infographics and character consistency, with 1K images at half the API price

Why it matters

Key points

Google released Nano Banana 2.1, its image generation and editing model, in the early hours of October 7 Beijing time (October 6 US time). It upgrades February's Nano Banana 2 and is now built on Gemini 3.6 Flash. It is rolling out across the Gemini app, AI Mode in Search, Google AI Studio, Flow, Stitch, Google Ads and Gemini Enterprise Platform, and developers can call gemini-nano-banana-2.1 in the Gemini API, where it launched as generally available with no preview stage.

Google highlights three areas of improvement: visual quality, mask-based editing, and keeping people and objects consistent across images. It accepts up to 14 reference images (consistency for up to 4 characters and 10 objects), renders text and infographic layouts more accurately, fixes tiling artifacts on extreme wide and tall aspect ratios at 2K and 4K, can ground images in Google Search results, and offers selectable thinking levels.

It is also much cheaper. On the Gemini API price list, a 1K image costs about $0.034, 2K about $0.05 and 4K about $0.11, versus about $0.067, $0.10 and $0.15 for Nano Banana 2, so 1K and 2K images are roughly half price and 4K about a quarter cheaper. Input pricing rose from $0.50 to $1.50 per million tokens, though, so workloads heavy on reference images save less. The older model was marked deprecated the same day, with no shutdown date announced yet.

Verifiable

Google DeepMind model card: https://deepmind.google/models/model-cards/nano-banana-2-1/ , published October 6, 2026. The Gemini API changelog records the general availability and the deprecation on the same date: https://ai.google.dev/gemini-api/docs/changelog . Model page: https://ai.google.dev/gemini-api/docs/models/gemini-nano-banana-2.1 ; pricing: https://ai.google.dev/gemini-api/docs/pricing . Google's announcement post: https://x.com/Google/status/2107501209154204148 .

Human side-by-side Elo scores from the model card: on overall text-to-image preference, Nano Banana 2.1 scores 1050 with thinking and 1015 without, versus 990 for Nano Banana 2 and 935 for Nano Banana Pro; on multi-character consistency the figures are 1106, 1068, 978 and 1011. Infographic factuality, measured by an autorater, rises from 0.179 for Nano Banana 2 to 0.521. These are Google's own evaluations and have not yet been checked by third-party leaderboards. The card also lists known weaknesses: small text is often blurry at 1K, and long paragraphs and character consistency are still imperfect. Its frontier safety assessment found no tracked or critical capability levels reached.

Our take

The notable thing here is not prettier pictures but that image models are now being tuned to carry information. A near-threefold jump in infographic factuality, better text rendering and search grounding all point to one use: having AI produce charts and explanatory graphics that go straight into reports, ads and web pages, not just illustrations. That matches OpenAI's recent move to have GPT-6 generate interactive interfaces inside answers; generated visuals are becoming part of the answer itself.

But read 0.521 the other way too: by Google's own scoring, nearly half the infographics still contain factual errors. Anyone using it for data graphics needs to check the numbers and labels by hand rather than trusting a professional-looking layout.

Price is the other story. Halving the 1K image price while deprecating the old model pushes developers to migrate wholesale and drags image generation pricing down to about three cents an image. What to watch next is independent evaluation on third-party image leaderboards, and whether OpenAI, ByteDance, Alibaba and others cut prices in response.