Nano Banana Pro Guide: What Gemini 3 Pro Image Really Costs

Most people evaluate an image model by looking at a handful of pictures. That tells you whether it can make a good image. It does not tell you what a thousand of them cost, how long each one takes, or which parts of your pipeline the API will refuse to do. For Nano Banana Pro those three questions matter more than the gallery.

Nano Banana Pro is Google's name for Gemini 3 Pro Image. Google announced it on 20 November 2025 as the next iteration of the original Nano Banana (Gemini 2.5 Flash Image). It takes text and images in and returns images and text out, it is built on Gemini 3 Pro, and Google's model page describes it as a reasoning-driven engine for image generation and editing.

What it is good at

Google's launch post leans on four capabilities, and they are a fair description of where the model differs from cheaper image generators:

• Legible text in images. Including multilingual text, plus translation or localisation of text already in an image.
• Composition from references. Google says it can blend up to 14 input images and keep the likeness of up to five people consistent across outputs.
• Grounded content. It can call Google Search to pull in real-world information, which is what makes infographics, weather charts and explainer diagrams feasible.
• Studio-style edits. Localised edits to a selected region, camera angle, focus, colour grading and lighting changes such as day to night.

Output comes in multiple aspect ratios at 1K by default, with 2K and 4K available. In the API you set the size through an image_size field, and the value must use an uppercase K: "4K" works, "4k" is rejected.

If your images are mostly decorative and contain no text, most of that list is capability you are paying for and not using. That is the main reason to read the next section before choosing it.

The bill is per image, so price the image

Google's pricing page lists Gemini 3 Pro Image at $2.00 per million input tokens and $12.00 per million text and thinking output tokens. Those are the figures you will see on aggregator listings, including OpenRouter at the time of writing. They are not the number that dominates an image workload.

Generated images are billed at $120 per million output tokens. A 1K or 2K image counts as 1,120 tokens and a 4K image as 2,000, which Google itself converts to roughly $0.134 per 1K/2K image and $0.24 per 4K image. Google prices each input image at about $0.0011, roughly a tenth of a cent. There is no free tier for this model on the Gemini API.

A worked example: 1,000 product renders at 2K cost about $134 in image output alone. The same job at 4K is about $240. If the renders can wait, the Batch tier halves those figures to roughly $0.067 and $0.12 per image, and batch API savings math covers when that latency trade is worth making.

Two practical consequences follow. Generate drafts at 1K and only re-render the winners at 4K. And count retries: an image you discard because the text came out wrong is billed the same as the one you keep.

Thinking is always on

Google's image-generation documentation says thinking cannot be disabled for its Gemini 3 image models. The model may produce up to two interim images to test a composition before the final one, and its thought steps can be read back through the API.

This is where much of its quality on complex prompts comes from, and it is also why it is slower and more expensive than a Flash-class image model. Thinking output is billed at the $12 per million text rate. For an interactive tool where a user waits on every render, measure end-to-end latency on your own prompts before committing, because no published figure will match your prompt lengths.

What the API will and will not do

This is the section that saves debugging time. Google's model page for gemini-3-pro-image lists image generation, Search grounding, thinking and the Batch API as supported. It lists function calling, structured outputs, context caching, code execution and URL context as not supported, and gives limits of 65,536 input tokens and 32,768 output tokens.

Some aggregator listings show tool calling or a larger context window for this model. Trust the vendor page. If your pipeline needs a structured description of the image alongside it, generate that with a text model in a separate call rather than expecting JSON from the image model. Best model for structured output covers that half of the job.

Every generated image carries Google's invisible SynthID watermark. Google also says images from free and Google AI Pro users in the Gemini app keep a visible watermark, while images made in Google AI Studio do not. If provenance matters to your customers, that is a feature; if you assumed images were unmarked, it is something to tell them.

Preview and stable model IDs

The model launched as gemini-3-pro-image-preview. The stable gemini-3-pro-image reached general availability on 28 May 2026, and Google shut the preview down on 25 June 2026, although some catalogues still list both. If code still names the preview ID on the Gemini API, it is failing. Switch deliberately and re-run your visual checks rather than assuming the outputs are identical. Pinning model versions explains why aliases break production quietly.

When a Flash image model is enough

Google's own documentation calls the Pro model the premium choice for the most complex visual tasks, and positions its Flash image models beside it. Nano Banana 2 (Gemini 3.1 Flash Image) offers a 512px option the Pro model does not, and the docs list Image Search grounding for the Flash models but not for Pro.

A reasonable rule: use Pro when the image contains text a reader must be able to read, when you need the same people across a series, or when the content has to be factually grounded. Use a Flash image model for thumbnails, backgrounds, mood boards and anything generated in bulk where a weaker image costs you nothing. The Gemini 3.1 Flash Image guide covers that side of the trade.

A test that tells you something

Do not judge it on five showcase prompts. Build twenty from real requests and score them the way your users would:

1. Include at least five prompts that require exact text, such as a product name, a price or a date, and count spelling errors.
2. Include a three-image series featuring the same person and check whether a stranger would say it is the same person.
3. Record wall-clock time per image and the number of renders you discarded.
4. Divide total spend by usable images, not generated ones.

Run the same set through a Flash image model. If the cheaper model's usable-image rate is close, the price gap decides it. How to benchmark LLMs on your own work has the general method.