Nano Banana 2.1 Halves Google's Image Price, Except Where It Doesn't

Google released Nano Banana 2.1 on October 6, an update to its Flash-tier image model that the company built on Gemini 3.6 Flash, according to the Google DeepMind model card. The headline is price. On Google's Gemini API pricing page, image output now costs $30 per million tokens, half the $60 charged for Nano Banana 2. A standard 1K image drops from $0.067 to $0.0336.

The fine print is more interesting than the headline. The halving is real at 1K and 2K. At 4K it is closer to a quarter. And the cost of everything a developer sends into the model went up.

The price table, line by line

According to Google's pricing page, Nano Banana 2.1 bills a 1K image at $0.0336, a 2K image at $0.0504 and a 4K image at $0.113 on the standard paid tier. Batch requests cost half that: $0.0168, $0.0252 and $0.0567. Nano Banana 2, which Google lists as Gemini 3.1 Flash Image, charged $0.067, $0.101 and $0.151 for the same three resolutions.

The 4K figure is where reports diverge. Android Headlines and several other outlets listed the new 4K price as $0.0756, which would be an exact halving. Google's own footnote explains why it is not. An output image consumes 1,120 tokens at 1K, 1,680 at 2K and 3,780 at 4K, and the per-image price is that count multiplied by the $30 rate. At 4K, that comes to about $0.113, a cut of roughly 25% from Nano Banana 2.

The input side moved the other way. Google's pricing page lists Nano Banana 2.1 input at $1.50 per million tokens for text, images and video, against $0.50 for Nano Banana 2. Text and thinking output rose from $3.00 to $7.50 per million tokens. There is no free tier for the new model.

For most single-image calls the trade still favors the new model. At 1K, the output saving is about 3.3 cents per image, and an extra dollar per million input tokens only cancels that out once a request carries more than roughly 33,000 input tokens. That threshold is not far-fetched for the workloads Google is marketing hardest, which lean on stacks of reference images. Teams running heavy multi-reference jobs should price them out rather than assume the 50% figure.

What the model does differently

Google's developer documentation describes Nano Banana 2.1 as keeping Flash-level speed and cost while improving visual quality, prompt adherence, multi-turn character consistency and text rendering. The documentation lists multi-image fusion with up to 14 reference images, character consistency for up to four characters and object fidelity for up to 10 objects. It also says Google fixed tiling artifacts on very wide and very tall aspect ratios, such as 1:8 and 8:1, at 2K and 4K.

Developers get a configurable thinking level, with minimal, medium and high settings, plus grounding through Google Web Search and Google Image Search, so the model can check real-world details before it renders. The documentation lists the model as gemini-nano-banana-2.1 with Batch API support. It does not support function calling, structured outputs or caching.

The DeepMind model card puts numbers on the quality claims. In human side-by-side tests with thinking enabled, Nano Banana 2.1 scored an overall Elo of 1050, against 990 for Nano Banana 2 and 935 for Nano Banana Pro, Google's larger image model. The biggest gap was in multi-character consistency, where 2.1 scored 1106. On an automated infographic factuality score, 2.1 posted 0.521, against 0.179 for Nano Banana 2. These are Google's evaluations, run on Google's methodology.

The card is also candid about limits. It says small or blurry text, long paragraphs and multi-page text still render poorly, that character consistency is not always perfect, and that the model occasionally confuses left and right.

Where it ships

The model card lists the Gemini app, Google AI Studio, the Gemini API, AI Mode in Search, Google Ads, Flow and Stitch as surfaces, with enterprise access through the Gemini Enterprise Agent Platform. Android Headlines reported that the rollout is sequential, so not every surface will have it on day one.

For developers already on Nano Banana 2, the decision has been made for them. The Gemini API changelog marks gemini-3.1-flash-image as deprecated and tells developers to migrate to the new model.

Building on the cheap tier

Flux wrote in July about Nano Banana 2 Lite, Google's attempt to own the low-cost end of image generation. Nano Banana 2.1 pushes the same strategy one tier up. Google is now selling a Flash-class model that, on its own benchmarks, beats both the model it replaces and its Pro model, at half the output price for the resolutions most production pipelines use.

The pattern is a familiar one in model pricing: cut the number everyone quotes, raise the numbers fewer people check. Image output is the line item that shows up in comparison charts, so it fell. Input, which grows with every reference image and every edit turn, rose. For an ad team generating thousands of 1K variants from short prompts, 2.1 is close to a straight 50% saving. For a studio feeding 14 reference images into 4K renders, the math is closer than the launch coverage suggests.

Either way, the price floor for a production-grade AI image keeps falling. At $0.0168 per batch 1K image, a million images now cost less than $17,000 in output fees on Google's own pricing.