Nano Banana Pro pricing: what one image actually costs

Three routes to the same model, priced three different ways

Nano Banana Pro is Google's gemini-3-pro-image. There is one model and three ways to be billed for it, and almost every disagreement you will find about "the price" is really two people standing on different routes:

• The Gemini app, on a subscription. You pay a monthly fee and generate inside Google's own product. Nothing is metered per image, and — this is the part that decides the whole comparison — Google's subscription page publishes no image quota at all. A monthly fee with no stated allowance cannot be turned into a price per image.
• The Gemini API, metered. You bring a Google API key and pay per token. Google publishes the per-image equivalents itself, which makes this the only route with a first-party number: $0.134 for a 1K or 2K image, $0.24 for 4K, halved in batch mode to $0.067 and $0.12. There is no free tier on this model.
• Somebody reselling it. Marketplaces and API brokers buy at volume and resell, either per image or in credits. This is where the price spread lives, and where nearly every page ranking for this query sits.

Clize is on the third route and says so plainly. What differs is the shape of the price rather than its size: one flat rate, $0.05, for any image model on the hosted path, printed as a quote before a cent moves. To put your own batch size in, the image generation cost calculator compares eleven published prices in your browser. This page answers the model-specific question it does not: what does this model cost on each route.

Reading a credits plan in dollars per image

The two subscription sites that occupy this search result both sell credits, and neither states a dollar-per-image figure anywhere. One of them at least publishes its conversion table, which is enough to finish the arithmetic it left undone. nano-banana-pro.io answers "how are credits calculated?" in its own FAQ: Nano Banana Pro costs 12 credits at 1K, 15 at 2K and 25 at 4K. Every plan then becomes divisible (all figures read 4 September 2026):

Same model, same site, $0.21 to $0.60 an image depending on which box you tick — a spread of nearly 3×, invisible until you divide. At 4K the divisor is 25 instead of 12, so every row roughly doubles.

nanobanana.io states an image count beside each plan instead, which comes to the same thing: Basic is $118.80 a year for 15,000 credits, described as roughly 750 images — about $0.16 each. Pro, $286.80 a year for roughly 1,800 images, lands in the same place; Enterprise, $766.80 for roughly 6,000, comes down to about $0.13.

Two things fall out that the plan pages will not tell you. A subscription is a cap, not a rate: Basic cannot produce 400 images in a month whatever you pay, because its whole year is 750. And both sites' cheapest per-image rates are their most expensive commitments — the $0.21 row is an $839.88 annual payment made in advance.

The one tool-shaped result here deserves a sentence. pricepertoken.com keeps a per-provider price page for this exact model, and on 4 September 2026 it reads, verbatim: "No provider pricing available for this model." A dedicated price comparison site has no price for it. That is the size of the gap.

Every published rate side by side, with the date it was read

One table, one resolution tier (1K/2K, the default), every figure carrying where it came from. Nothing here is averaged or estimated; where a route does not publish a per-image number, the cell says so instead of guessing.

Two footnotes on the official row, since it is the one everyone quotes. The $0.134 is the image output component only: input is billed separately at $2.00 per million tokens (Google puts that at about $0.0011 per input image), and Nano Banana Pro is a thinking model whose text and thinking output runs $12.00 per million on top. A real API bill therefore sits a little above $0.134, never below. And batch mode buys its discount with latency — not the same product as a synchronous call.

What the flat $0.05 is, and is not: a resale rate on a date, which currently happens to sit under Google's own standard rate for this model. Not a permanent guarantee, and not a claim to be the cheapest anything — kie.ai's tiered $0.12 may well be the better deal for 4K-heavy work. Flat buys forecastability, not a floor: 400 images is $20.00 before the first prompt is written, and still $20.00 after you switch models mid-batch.

The line items each route bills differently

Comparing headline rates is the easy half. What decides a real invoice is how each route treats the generations you did not want, the reference images you attached and the variants you asked for at once. On Clize, measured on 4 September 2026 by running the quote command and reading what came back:

Reference images. A product photo, a character sheet, three views of the same object — on a per-token API these are metered input, on a credits plan often a tier feature. Here they change no number. The only limit is the model's: nano-banana-2 takes 14, gpt-image-2 takes 16, and going over is refused before anything is charged:

$ clize gen image "the same bottle, new background" --model nano-banana-2 --ref r1.png,…,r15.png
❌ [400] At most 14 reference images (upstream limit of model nano-banana-2;
   gpt-image-2 takes 16, nano-banana models 14; got 15).

Retries. Nobody keeps the first generation. Every route above bills the throwaways at the keepers' rate, and so does this one — flat is not a keep-rate guarantee. What is not billed is a generation that fails, which is a different event from one you dislike: provider errors and empty results are refunded in full, automatically, as a credit on your balance.

Several at once. One asymmetry to know before planning a batch: --n only means anything on nano-banana-2. The gpt-image-2 path runs one image per job and refuses with a reason rather than silently making one:

$ clize gen image "a red bicycle" --n 4
❌ [400] gpt-image-2 uses the async path and makes 1 image per job;
   run several jobs for more, or use --model nano-banana-2.

One naming detail, since this page is about a specific model. On Clize's hosted path --model nano-banana-2 resolves upstream to nano-banana-pro — it is a single line in the provider's alias table. If you want Nano Banana Pro through Clize, that is the flag.

The quote comes before the charge, every time

Everything above is a number on somebody's marketing page, including ours. Here is the part you can verify in thirty seconds without spending anything, which is the only reason to believe any of it. Run the generation command without --confirm:

$ clize gen image "a red bicycle leaning on a wall" --model nano-banana-2
{
  "quote": {
    "model": "nano-banana-2",
    "n": 1,
    "size": "1024x1024",
    "estUsd": 0.05,
    "currency": "USD"
  },
  "message": "📋 Quote about $0.05 (1 image(s)). Add --confirm to generate
   (confirm: true in MCP; charged to your clize balance, refused if it is insufficient)."
}

That is the actual output, copied from a terminal on 4 September 2026. No image was generated and no money moved, because the flag was not there. Adding --confirm is what generates and charges; there is no pre-approved allowance that skips it and no ceiling you can raise to make it go away — every generation is confirmed on its own, every time. That is a deliberate constraint rather than a missing feature, and the reason this is safe to hand to an agent running unattended.

Afterwards the ledger is readable rather than inferred. clize gen list --modality image prints every asset with model, prompt, state and cost, and clize balance groups the month by category — on the account used to check this page it reads "usd": 0.9, "n": 18 for media, which is eighteen images at five cents and nothing else. If a page tells you a price, ask it for the row that proves it.

Two neighbours run on the same gate: video at $0.48 a clip, music at $0.12 a track. The mechanics from a terminal — flags, output contract, long jobs — are in generating images from the command line.

What a flat rate does not cover

A pricing page that only lists reasons to buy is a brochure. These are the edges of the flat rate, as of September 2026, and any of them may be the reason it is wrong for you.

• No resolution tiers. Google prices 1K/2K and 4K separately; Clize publishes no 4K SKU. If your work depends on that path specifically, price it where it is named.
• No free tier and no trial credits. Google's API has none for this model either, but several subscription sites hand out starter credits. If the real question is "can I try it for nothing", the answer here is no.
• No user-settable spending cap. The gate is per generation: run the command, read the quote, add --confirm. There is no monthly allowance to configure, so the protection against a runaway loop is the mandatory flag on every charge, not a ceiling that catches it afterwards.
• Prepaid balance, not an invoice. A command is refused when the balance will not cover the quote rather than overdrawing it.
• One price, so no volume discount. Six thousand images cost six thousand times five cents. The annual-commitment routes get cheaper as you scale into them; this one does not.
• Prices move. Every third-party figure here carries its reading date. Believe the vendor's own page over ours, and tell us when a row has drifted.

What is left is a narrow claim: if you want one forecastable number, no Google API key to obtain, a quote before any charge and nothing billed for a failure, the flat shape fits. If you want the lowest unit cost at 4K and do not mind managing a key and a rate card, it does not. Both are honest answers to "what does Nano Banana Pro cost" — they belong to different questions.

$ npm i -g @clize/clize && clize login
$ clize gen image "a red bicycle leaning on a wall" --model nano-banana-2