Nano Banana 2.1 Halved the Price of an AI Image. Your References Got More Expensive.
Google's new image model costs half as much per picture as the one it replaces. Every launch story led with that. Fewer mentioned the other line on the same pricing page: the cost of the material you hand the model, your reference photos and your prompt, went up threefold.
Both numbers are real. Which one matters depends on how you work. If you type a sentence and take what comes back, Nano Banana 2.1 costs half what its predecessor did. If you build scenes from a stack of character sheets and product shots, which is exactly what this model was built for, the discount is smaller than the headline. Still a discount. Just not half.
That gap is worth understanding, because it points to the right way to use this model. Draft cheap, commit late, and treat references as a budget line.
What Nano Banana 2.1 is for
Google released Nano Banana 2.1 on October 6. The Gemini API page lists it as a stable model, and Google's deprecations page names it the recommended replacement for seven older image models, including Nano Banana 2 and the retired Imagen 4 family. In practice, this is the model Google wants you on for everyday image work.
The feature that matters most to storyboard artists, comic makers and product designers is the reference count. You can hand it up to 14 images in one request, and Google says it keeps up to four characters and ten objects recognizable in the result. Four characters is a two-shot with a background couple. Ten objects is a product line on a shelf.
It also renders at 1K, 2K or 4K, handles panoramic frames up to 8:1 (Google says the seams that used to show up in wide frames at 2K and 4K are fixed), and lets you pick how much the model "thinks" before it draws. There are three settings, minimal, medium and high. Think of it as the difference between a quick sketch and a planned layout. The extra planning pays off most on posters, infographics and anything with words in it.
You can reach it in the Gemini app, Google AI Studio, Flow and Stitch, and since October 7 as a node inside ComfyUI (version 0.39.1), the free node-based tool many AI artists build their pipelines in.
The two prices, side by side
Here is what Google's pricing page lists for the paid API, per finished image:
| Size | Nano Banana 2.1 | Nano Banana 2 |
|---|---|---|
| 1K | $0.0336 | $0.067 |
| 2K | $0.0504 | $0.101 |
| 4K | $0.113 | $0.151 |
At 1K and 2K that is half price. At 4K it is about a quarter off, not half. Batch jobs, where you submit a pile of requests and wait for them, cost half again across the board.
The input side runs the other way. Text, images and video you send in cost $1.50 per million tokens on Nano Banana 2.1, against $0.50 on Nano Banana 2. A token here is just the unit Google uses to meter what you send; a short prompt is a few dozen of them, and an image is a lot more.
How many more? Google's media resolution page lists 1,120 tokens per input image at the default setting for its Gemini 3 models. Google does not state the figure for Nano Banana 2.1 specifically, so take what follows as an estimate. If each reference costs about that much, a full stack of 14 references adds a little over two cents to every request on 2.1, against under a cent on the old model.
Run the numbers for a 1K storyboard frame with all 14 references loaded and the old model comes to roughly seven and a half cents, the new one to just under six. Cheaper, by about a quarter. Half price only shows up when you send few or no references.
None of this is a scandal. It is a pricing choice that rewards the habit most people should have anyway.
Why the model changes the way you should draft
For most of the last two years, the sensible way to work with a paid image model was to generate at the size you needed, because rerunning at a bigger size often changed the picture enough to lose what you liked. That habit is expensive at 4K.
Nano Banana 2.1 makes a different rhythm affordable. A hundred 1K drafts cost about $3.36 in raw output price. You can rough out an entire 30-panel storyboard three times over, throw away most of it, and still spend only a few dollars.
The thinking setting adds a second dial. Minimal thinking is the right choice while you are exploring composition and cast. High thinking is for the frame you will actually present, especially if it carries a title, a caption or a chart.
So the work splits into two passes. A cheap, loose exploration pass at 1K with the references you actually need. Then a slow, careful final pass at the size and thinking level the job deserves.
Put this into practice
The lowest-friction way in is the Gemini app or AI Studio. Google's own pages do not list a free API tier, and I could not find Google's published daily limits for 2.1 in the app; one outlet reports it is free to try there within each plan's existing image allowance. Start there before you spend anything.
- Build a reference sheet first. Pick the characters and props that must stay consistent. One clear, well-lit image per character works better than five messy ones, and it costs less.
- Name things in plain words. Describe the scene the way you labeled the references: "the woman in the red coat from image one hands the brass key from image three to the boy from image two." The model reads every reference alongside your words, so a sentence that points at each one gives it less to guess.
- Draft at 1K with minimal thinking. Generate several variations per frame. Judge composition, staging and likeness, nothing else.
- Trim the reference stack for each frame. If a shot only shows two characters, send two references, not all fourteen. That is the habit the new pricing rewards.
- Commit at the end. Rerun only the keepers at 2K or 4K with thinking set to high, and check every letter of any text in the frame.
- Batch anything that can wait. If you are producing a set overnight, the batch price halves the bill again.
If you already build in ComfyUI, update to version 0.39.1 and the Nano Banana 2.1 node handles both making and editing images, with the same 14-reference limit and the three thinking settings. Remember it is a paid connection to Google, not a model running on your own card, and ComfyUI's new --disable-partner-nodes launch setting switches it off.
Where it falls short
Google's model card is unusually frank, and the limits it lists are the ones you will hit first.
Small text still fails. The card says small text, long paragraphs and full-page text often render poorly, and that small text is often blurry at 1K. So judge layout in your 1K drafts, not lettering. A poster headline is fine. A menu is a gamble. Plan to set body copy yourself in a layout tool.
Consistency is good, not guaranteed. The card says character consistency between input and output "is not always perfect." Across a 30-frame storyboard, expect a few faces to drift and budget time to regenerate them.
Mask and doodle edits are partial. If you draw on an image to tell the model where to change things, the card says it follows those instructions only partly. Precise retouching still belongs in your editor.
It mixes up left and right. The card mentions occasional left-right confusion and rare cases where it keeps the pose from your input instead of the one you asked for. Blocking-heavy shots need a check.
The benchmark is Google's. The headline preference score, 1,050 against 990 for Nano Banana 2, comes from Google's own side-by-side tests of the "thinking" version. Nobody independent had published comparisons when this was written.
The ground moves under hosted models. The same deprecations page that crowns 2.1 lists seven models it replaces, and ComfyUI's changelog is removing another company's video model, Luma Ray 2, on October 24. A pipeline built on a hosted model runs on someone else's calendar. Keep your prompts and reference sheets in your own folders, so a model swap costs you an afternoon of retesting rather than a lost project.
The question to ask before your next job
The real shift here is small and practical. Consistency across a cast used to be the expensive part of AI image work, and it just got cheaper, as long as you are deliberate about what you feed the model.
So before your next storyboard, try this. Make one frame twice at 1K: once with every reference you have, once with only the two or three the shot needs. Compare the likeness and compare the bill. If the trimmed version holds up, you have just learned the habit that makes this model worth switching to. If it doesn't, you have learned exactly how much consistency costs you, which is a number worth knowing before a client asks for thirty more.
Medium metadata
Title: Nano Banana 2.1 Halved the Price of an AI Image. Your References Got More Expensive.
Subtitle: Google's new image model keeps up to four characters and ten objects consistent for about three cents a draft. Here is what the half-price headline leaves out, and the draft-cheap, commit-late habit that makes it pay off.
Tags: AI Art, Storyboarding, Generative AI, Design, Google Gemini
Estimated read time: 7 minutes