Google Launches Nano Banana 2.1: Better Than Its Predecessor at Half the Price

What it claims

Google released Nano Banana 2.1 on Oct. 6, 2026. It is an image generation and editing model that succeeds Nano Banana 2 (February, built on Gemini 3.1 Flash Image). Google names three upgrades:

• Visual design.
• Mask-based editing. You mark one region and only that area changes.
• Subject consistency. A person or object keeps its identity across several edits.

The developer API costs about half of what Nano Banana 2 did, at every listed resolution.

Evidence

All figures are Google's own. The page states that independent testing was not available at launch.

• Infographic factuality: 0.521 is reported "versus 0.179". The page does not say which predecessor scored 0.179, so it is not assigned to a column above.
• ELO: this is a relative preference rating with no ceiling. The page gives no evaluator pool, prompt set or match count.

How to use it

Where it is available: Google AI Studio, Flow and Stitch for developers, plus the Gemini app, Search AI Mode and Google Ads.

Inputs and outputs:

• Up to 14 reference images.
• Tracks up to 4 characters and 10 objects.
• Output up to 4K, with aspect ratios as wide as 8:1.

Controls:

• Thinking time is set per request, from "minimal" to "high".
• Results can be grounded in Google Search and Google Image Search.

API pricing per image:

• Batch (non-realtime) jobs get a further 50% off.
• That makes roughly $33.60 per 1,000 images at standard 1K, against $67 before. Batch would bring this to about $16.80, assuming the discount stacks as described.

The page gives no model ID string, no request schema, and no pricing for the thinking levels.

Limits

• Unverified quality claims. Every quality number is self-reported, and no outside benchmark existed at publication.
• Unknown benchmark setup. The ELO setup and the factuality benchmark are not described.
• Ambiguous factuality baseline. The 0.179 figure is not tied to a named model.
• Thinking-level costs not stated. The page does not say whether higher thinking levels or search grounding add latency or cost. Check this before budgeting from the per-image prices.
• Editorial wording. The article's "extremely good" framing is the writer's own, not a measured result.