Nano Banana 2.1's half-price images save money only on input-light jobs

The per-image price comes from a token bill. According to a guide published on dev.to, Google's Gemini API pricing page listed image output tokens at $30 per million on October 7, down from $60 ^[5]. If the per-image price is all output tokens, a 1K image works out to 1,120 tokens, a 2K image to 1,680 and a 4K image to 2,520 ^[19]. The saving is a rate cut. The per-token rate and the per-image price both fell by half, so the implied token count per image is unchanged from Nano Banana 2 ^[20].

Input went the other way. The same guide says 2.1 charges three times as much for input tokens ^[10]. Call a request's input spend on Nano Banana 2 I, and its output spend O. On 2.1 the same request costs 3I + O/2. The new model is cheaper only while I stays below a quarter of O ^[21]. At 1K, where Nano Banana 2 works out to $0.0672 an image ^[18], the ceiling is $0.0168 of input spend per image ^[22]. The calculation assumes both models turn the same prompt and references into the same number of tokens.

Editing is where input grows. A 2.1 request can carry up to 14 reference images, and keeping a character or product consistent across edits means chaining them with previous_interaction_id ^[15]. The guide's own recommended workflow ends with logging input token consumption, especially when several references are attached ^[16]. Text prompts are short, so I'd expect prompt-only generation to stay well under the line. A pipeline that attaches most of the 14 allowed references to every edit is the one to meter.

The quality case rests on Google's word. The only launch statement was a brief @GoogleAIStudio post saying the model outperforms Google's previous models in every respect ^[4]. Every respect is a lot of ground for a post that short. Google's documentation lists accurate text rendering among the 2.1 improvements ^[7], and the launch post cites better visual design, mask-based editing, subject consistency and more natural-looking images ^[12]. The dev.to guide treats the text gain as a Google claim, not a measured result ^[23]. Its suggested test set before moving production traffic is menus, product labels, posters, slides, interface mockups and infographics ^[17].

Grounding has a limit that matters for some products. The model can ground generation in Google Image Search and web search, but it cannot use real images of people pulled from web search ^[9]. Google keeps Nano Banana 2 as the fast option, Nano Banana 2 Lite in the economy slot and Nano Banana Pro for professional assets, with the original Nano Banana now legacy ^[14].

In my view, 2.1 is the right default for prompt-only generation at 1K and above, where the halved output rate passes straight through to the bill ^[20]. For reference-heavy editing I would run both models on a sample of real requests and compare input token counts before moving traffic ^[21].