Nano Banana 2 vs Nano Banana Pro: Which Gemini Image Model for Ads

One Line of Code Decides What Your Ad Images Cost

A growth marketer at a skincare brand is wiring up a static-ad pipeline: forty variants a week, 4:5 for the feed, 9:16 for Stories. The prompt gets all the attention. The line that sets the bill sits one field above it, model. On Google's Gemini API pricing page, last updated September 16, 2026, a 1K image costs $0.067 on Nano Banana 2 and $0.134 on Nano Banana Pro at the Standard paid tier.

Both of those are Nano Banana. As of September 2026, Google's image-generation guide (last updated September 17, 2026) says the name "refers to four distinct models available in the Gemini API": Nano Banana 2 Lite, Nano Banana 2, Nano Banana Pro and the original Nano Banana. The original is scheduled to shut down on October 2, 2026, and Lite is the speed-and-cost tier we covered in our Nano Banana 2 Lite breakdown. That leaves two full-size models for ad work: Nano Banana 2 (gemini-3.1-flash-image) and Nano Banana Pro (gemini-3-pro-image).

This post sets them side by side on what an ad team actually pays for: price per image at each resolution, how many reference images each one holds, grounding, and how long each id is likely to stay live. Every Google figure is dated to the docs page it came from. The ratios between them are our own arithmetic, and we say so each time.

The quick answer: which Nano Banana to call for each ad job

Google's own model-selection note says Nano Banana 2 "should be your go-to image generation model" and describes Nano Banana Pro as "designed for professional asset production and complex instructions". That is the vendor's positioning, not a test result. The documented specs, though, split the two in a way an ad team can act on.

Use Nano Banana 2 when

• You are drafting concepts or thumbnails, where its 512px tier is the cheapest render in the pair
• The brief carries many product shots, since Google's table lets it hold up to 10 object references
• A placement needs a banner shape such as 8:1 or 1:4
• You want grounding in Google Image Search results, not only web text
• You want a thinking setting you can turn up or down per job

Use Nano Banana Pro when

• The scene has to keep five recurring people consistent (Google's table gives Pro up to 5 character references, Nano Banana 2 up to 4)
• A brand look has to carry across a set, since only Pro's column lists style references
• The final is 2K, where Pro charges the same as its own 1K and the gap to Nano Banana 2 is narrowest

Default: Nano Banana 2. Move a brief to Pro when its reference set asks for Pro.

Where the two sit in Google's lineup

Google does not sell Nano Banana as one model with a quality slider. It documents four models with four different jobs, and the two in this comparison are the middle of that range.

Four models, as of September 2026

The image-generation guide lists them in this order:

• Nano Banana 2 Lite (gemini-3.1-flash-lite-image): "our fastest and cheapest Gemini image model", which Google says is "not optimized for multiple reference inputs or multi-turn sequential editing"
• Nano Banana 2 (gemini-3.1-flash-image): "the most versatile model, generalist workhorse model for all tasks"
• Nano Banana Pro (gemini-3-pro-image): "the premium choice for the most complex visual tasks"
• Nano Banana (gemini-2.5-flash-image): "the legacy pioneer of the Nano Banana series", which Google now tells customers to move off

The longer descriptions overlap more than the labels suggest. Google says Nano Banana 2 "balances speed with state-of-the-art 4K generation, world knowledge, and reliable text rendering" and is "excelling at multiple reference image processing and consistency". For Pro it promises "the highest level of world knowledge, advanced localization, accurate brand consistency, and precision creative control".

Both descriptions claim text, consistency and world knowledge. Neither is a benchmark, and this post does not cite one. Treat quality as something you measure on your own product, and treat the rest of this comparison as what the docs let you plan around.

How long each id will be around

Google's deprecations page, last updated September 17, 2026, lists gemini-3.1-flash-image and gemini-3-pro-image as released on May 28, 2026, and both rows read "No shutdown date announced". Neither model has a published end date.

The preview ids are a different story. Google introduced Nano Banana 2 on February 26, 2026, and the same table dates gemini-3.1-flash-image-preview to that day. Pro's preview id, gemini-3-pro-image-preview, dates to November 20, 2025. The table gives both previews a shutdown date of June 25, 2026 and names the GA id as each one's replacement. If an old script still ends in -preview, that suffix is the first thing to fix.

The original's clock makes this a migration decision

The pricing page carries a warning on the original model: gemini-2.5-flash-image "is deprecated and will be shut down on October 2, 2026", and Google tells users to "migrate to Gemini 3.1 Flash Image or Gemini 3.1 Flash Lite Image". The deprecations page adds a caveat that cuts both ways: its dates are "the earliest possible dates on which a model might be retired".

Our read: a team still calling the original has less than two weeks to choose between Lite, Nano Banana 2 and Pro. Google's warning names the first two. Pro is a valid destination too, and for some briefs the right one, which is what the rest of this post is for.

Price per image at every resolution

Every price below is from the Gemini API pricing page, last updated September 16, 2026: the Standard paid tier, output images. The right-hand column is our division, not a Google figure.

Output size	Nano Banana 2	Nano Banana Pro	Pro vs Nano Banana 2 (our math)
0.5K (512px)	$0.045	Not listed	n/a
1K	$0.067	$0.134	2.0x
2K	$0.101	$0.134	about 1.33x
4K	$0.151	$0.24	about 1.59x

The 0.5K row exists because, in Google's words, "Gemini 3.1 Flash Image adds the smaller 512px (0.5K) resolution" to the 1K, 2K and 4K sizes both models share. By our arithmetic, three Nano Banana 2 drafts at 512px cost about the same as one Pro render at 1K.

Why the gap shrinks at 2K

Google prices image output by the token, and the token math explains the odd shape of that table. Nano Banana 2 output is "priced at $60 per 1,000,000 tokens" and Pro output at "$120 per 1,000,000 tokens", so at an equal token count Pro costs exactly twice as much.

The token counts are where the two diverge:

• Nano Banana 2 uses 747 tokens at 0.5K, 1,120 at 1K, 1,680 at 2K and 2,520 at 4K, so every step up in size costs more
• Nano Banana Pro bills images "from 1024x1024px (1K) and up to 2048x2048px (2K)" at the same 1,120 tokens, and 2,000 tokens at 4K

Two practical rules fall out of this, and they are ours, not Google's. On Pro, 1K saves nothing: 2K costs the same. On Nano Banana 2, 2K costs about 51% more than 1K by our arithmetic, so keep 1K or 512px for exploration and spend 2K on the frames you intend to ship.

Cost per keeper, not cost per render

Nobody ships every render. The number that matters is what one usable image costs, and that depends on how many attempts each model needs on your briefs.

The cost-per-keeper equation: cost per usable image = price per render x renders per keeper.

Take the skincare team's week: 40 finished statics at 2K.

• Pro, one render per keeper: 40 x $0.134 = $5.36
• Nano Banana 2, one render per keeper: 40 x $0.101 = $4.04
• Nano Banana 2, three renders for every two keepers: 60 x $0.101 = $6.06

That last line is the break-even at work. At 2K, Nano Banana 2 stays cheaper per keeper only while it needs fewer than about 1.33 renders for every one Pro needs. At 1K the threshold is 2.0, and at 4K it is about 1.59. We are not claiming either model needs more attempts. That ratio is yours to measure.

The more honest lesson is about scale. At 40 images a week, the whole difference is $1.32, so at that volume the choice is about keepers and review time, not the invoice. The price table starts to matter at catalog volume: by our arithmetic, a thousand 4K images a month comes to $151 on Nano Banana 2 and $240 on Pro. If you are unsure how many variants a week you actually need, our guide to creative volume works that number out from your conversion data.

References: the spec that decides product ads

For a product brand, the most important table on Google's image-generation page is not the price table. It is the one under "Use up to 14 reference images", which splits those 14 slots differently for each model.

Objects versus characters

Google's table gives each model a different mix:

• Nano Banana 2: "Up to 10 images of objects with high-fidelity to include in the final image" and "Up to 4 images of characters to maintain character consistency"
• Nano Banana Pro: "Up to 6 images of objects with high-fidelity to include in the final image", "Up to 5 images of characters to maintain character consistency" and "Up to 3 images to be used as style references"

Our reading: the counterintuitive part is that the cheaper model is documented for more product references. For a brand whose ads show a range (a kit, a bundle, a shade line), that matters more than the price table. Pro's extra slots go to people and to style, which is where campaign work tends to need them.

Two numbers for Pro, so plan for the lower one

The same page states the limits a second time, in a list of limitations further down. There, Google says gemini-3-pro-image "supports 5 images with high fidelity, and up to 14 images in total", and gemini-3.1-flash-image "supports character resemblance of up to 4 characters and the fidelity of up to 10 objects in a single workflow".

The Nano Banana 2 figures match the table. Pro's do not quite: 5 high-fidelity images in the note against 6 object images in the table. Until Google reconciles the two, build a Pro reference set that works at five.

Mapping a real reference set

Picture a supplement brand's launch static. The reference set is five object images (bottle front, bottle back, a gummy close-up, the box, the logo lockup), one recurring creator, and two past campaign images that carry the brand's look.

• On Nano Banana 2, the five objects and the one character fit. The two campaign images have no style row in its column, so the look, down to the brand's hex colors, has to come from the prompt.
• On Nano Banana Pro, everything fits the table: five of six object slots, one of five character slots, two of three style slots. Under the stricter note, the five products alone use the whole high-fidelity budget, so drop one or test whether the fifth holds up.

Now make it a seven-SKU bundle shot. Seven object references exceed Pro's table, while Nano Banana 2 still has three to spare. The same brand can land on different models for two briefs in the same week, and that is the correct outcome.

A slot count is a limit, not a quality score. Neither number says the label on your bottle comes out right, so check shape, color and label copy on every render; our Nano Banana Pro ad creative guide covers the read-back for text-heavy statics. When one recurring face matters more than the product, the method in keeping one AI actor consistent across ad variations is the place to start.

Grounding, thinking and banner shapes

The remaining differences are smaller, but each one can decide a specific kind of ad.

Search grounding, with images and a caveat

Google says Pro "features real-world grounding using Google Search", and that Nano Banana 2 "adds the integration of Google Image Search Grounding alongside Web Search". Grounding lets a prompt pull in real-time data, and Google's own examples are current weather maps, stock charts and recent events, instead of relying only on what the model already knows.

Two details matter before an ad team leans on it:

• People are excluded on Nano Banana 2. Google states that its grounding "does not support using real-world images of people from web search at this time".
• Grounding is billed separately from the image. After a monthly allowance shared across Gemini 3.x models, the pricing page charges $14 per 1,000 requests.

Our view: most product ads do not need grounding at all. Your own product photos are the references that matter, and they go in as images, not search results. Grounding earns its place in timely creative, such as a seasonal or event-led static.

A thinking setting on Nano Banana 2

Google describes Gemini 3 image models as "thinking models" whose reasoning "is enabled by default and cannot be disabled in the API". For Nano Banana 2 (and Lite), the docs add a control: "the default thinking_level is minimal", and "the supported levels are minimal and high", set "to balance quality and latency". Pro's description instead mentions "a default 'Thinking' process that refines composition prior to generation".

The interim "thought images" are, per Google, "visible in the backend but not charged". Our suggestion: leave Nano Banana 2 at minimal for drafting and try high on the briefs where it misses. Then check whether your renders-per-keeper number actually moves.

Banner shapes

Google's Nano Banana 2 launch post says "we have now added 4:1, 1:4, 8:1 and 1:8" to its aspect ratios, and the image-generation page's size tables bear that out. The Pro table on the same page lists ten ratios, running from 1:1 to 21:9.

By our arithmetic, a 728x90 leaderboard is roughly 8:1 and a 160x600 skyscraper is close to 1:4. If a display campaign needs those shapes, Nano Banana 2 is the model of the two whose size table covers them.

How to settle it on your own product

Google's docs tell you what each model is allowed to take in. Only a test tells you which one gives you keepers on your briefs. This one costs about seven dollars.

A two-model keeper test

1. Pick ten briefs from next week's real plan: some product-only, some with people, some with a headline in the image.
2. Render each brief three times at 2K on both models, 60 renders in all. By our arithmetic that is 30 x $0.101 plus 30 x $0.134, or $7.05.
3. Score every render keep or reject on the checks that sink ads: product accuracy, a legible headline, the right number of people, a crop that fits the placement.
4. Count renders per keeper for each model on each type of brief, then multiply by the price per render.
5. Apply the break-even from the price section: at 2K, Nano Banana 2 wins a brief type if its renders per keeper stay under about 1.33 times Pro's.

How to know it worked: every brief type ends with a named model and a cost per keeper. If you end with one overall favorite, you averaged away the answer.

Keep the comparison fair

• Same inputs: the same prompt, the same references (within the smaller model's slots) and the same aspect ratio
• Blind scoring: hide which model made which render until the scores are in
• Per-type verdicts: product bundles and people-led briefs may land on different models, and that is the point

A keeper is an image you would ship, not one that wins. The in-market question belongs to a proper split test, which our ad creative testing guide sets up.

What does not change with the pick

Google states that "all generated images include a SynthID watermark", so the watermark question is the same whichever model you pick; Google Ads' rules for AI-generated images cover what that means in the ad account. And if neither model passes your test, our Nano Banana alternatives list sorts the options outside Google's family by job.

How Novoads solves the Pro side of the choice

Nano Banana Pro is available in Novoads, in the same image picker and on the same credit balance as the video models an ad team uses next.

Pro in the picker, Nano Banana 2 outside it

• Nano Banana Pro in the Novoads image picker renders at 1K, 2K or 4K, at 2 credits for a 1K or 2K image and 4 credits for a 4K image.
• Nano Banana 2 is not offered. Novoads does not run Nano Banana 2, Lite or the original, so a brief the rule above sends to Nano Banana 2 belongs in Google's Gemini API.
• Plans are published on the pricing page, and credits work across images and video.

From the still to the ad

A static is where an angle gets decided, and it is rarely the whole campaign. In Novoads a finished still can be the starting image for a short clip from one of the video models, and a script plus an AI actor turns the same offer into a UGC-style video ad with voice, lip-sync and captions. Our guide to making product videos with AI walks through that handoff step by step.

The reference set picks the model

Google's price table makes the two models look like a ladder: Nano Banana 2 is cheaper at every size it shares with Pro. Its reference table makes them look like a fork. One side holds more products, the other holds more people and a style lock. Ad teams that pick by brand name will overpay on half their briefs and underfit the other half. Pick per brief, measure cost per keeper, and let the reference set make the call.