Is Nano Banana the same as Gemini?

Nano Banana is Google’s name for the image models in its Gemini family. Nano Banana 2 is Gemini 3.1 Flash Image, Nano Banana 2 Lite is Gemini 3.1 Flash Lite Image, and Nano Banana Pro is Gemini 3 Pro Image. The Gemini app uses Nano Banana 2 for everyone and offers Nano Banana Pro on paid plans.

What happened to Imagen?

Google shut down Imagen 4 on the Gemini API on August 17, 2026 and recommends Nano Banana 2 (gemini-3.1-flash-image) in its place.

Which model is cheapest for product photos?

Per render, GPT Image 2 at low quality ($0.006 in OpenRouter’s test), Meta’s Muse Image ($0.01 on OpenRouter) and Nano Banana 2 Lite ($0.034) are the cheapest well-known options in September 2026. The cheapest per usable photo depends on how many renders you throw away, so count keepers before you switch.

Why does GPT Image cost so little in one test and so much in another?

OpenAI prices GPT Image by tokens, and the quality setting changes how many tokens an image uses. OpenRouter found that leaving quality unset billed the same as low, $0.006 for a square image, while high cost $0.21. Set the quality on purpose whenever you compare prices.

Which AI image model is best for product photos?

Checked September 24, 20268 min read

Short answer

As of September 24, 2026, OpenAI’s GPT Image 2.5 models (Sunburst, then Flare) rank first on public leaderboards for both new images and edits. For product work, weigh price and fidelity too: Google’s Nano Banana 2 costs $0.067 per 1K image and Nano Banana Pro $0.134, against about $0.21 for GPT Image 2 at high quality. Rankings change month to month, so test two or three models on your own product.

The models worth testing, with prices

Nano Banana is Google’s name for the image models in its Gemini family; GPT Image is OpenAI’s. Prices are for one standard square image, from the maker’s own price list where it publishes one per image, as of September 24, 2026. Where a maker publishes no price per image, the figure comes from a published test by OpenRouter, a service that sells access to many AI models through one account, or from Artificial Analysis, which tracks model prices.

Model	Maker	Price per image	Good at	Watch for
GPT Image 2.5 Sunburst	OpenAI	Priced per token like GPT Image 2; about $0.21 at max quality	Edits that keep the product and change the scene; first on both Arena leaderboards	Slower than Flare
GPT Image 2.5 Flare	OpenAI	Same token prices as Sunburst	OpenAI’s default: higher quality than GPT Image 2 at half the latency	Text placement can still slip
GPT Image 2	OpenAI	$0.006 at low quality to $0.21 at high	Cheap drafts at low quality; transparent backgrounds	The quality setting changes the bill 35 times over
Nano Banana Pro	Google	$0.134 (1K or 2K), $0.24 (4K)	Text and world knowledge; up to 6 object and 5 character references	Ranks below OpenAI’s models on this month’s leaderboards
Nano Banana 2	Google	$0.067 (1K), $0.151 (4K)	Up to 10 object, 4 character and 3 style references; 512 px to 4K	Not in the top 5 on either Arena board
Nano Banana 2 Lite	Google	$0.034 (1K)	Fast, cheap tests; up to 14 object references	1K only; no character references
Seedream 5.0 Pro and Lite	ByteDance	$0.09 (Pro) and $0.035 (Lite), as billed by OpenRouter	Editing and text in many languages	ByteDance says fine text and editing consistency still need work
FLUX.2 Pro and Max	Black Forest Labs	From $0.03 (Pro) and $0.07 (Max)	A low price per image	The small FLUX.2 Klein model misspelled a label in OpenRouter’s test
MAI-Image-2.6	Microsoft	About $0.039 (Artificial Analysis)	Fourth on Arena’s board for new images	Check fidelity on your own product
Muse Image	Meta	$0.01 on OpenRouter	One of the lowest prices per image; seventh on Arena’s board for new images	Check fidelity on your own product

Two older Google models are leaving. Google shut down Imagen 4 on its Gemini API on August 17, 2026 and points users to Nano Banana 2, and the original Nano Banana (Gemini 2.5 Flash Image) shuts down on October 2, 2026. Midjourney’s V8.2 (July 2026) focused on aesthetics, and its V8.2 edit model takes up to four reference images.

What the leaderboards say this month

On an arena leaderboard, people see two images made from the same prompt by two hidden models and vote for the one they prefer; the votes add up to a score. Arena’s board for new images had about 6.5 million votes on September 24, 2026.

Rank	Arena, new images	Arena, edits	Artificial Analysis, new images
1	GPT Image 2.5 Sunburst	GPT Image 2.5 Sunburst	GPT Image 2.5 Sunburst (max)
2	GPT Image 2.5 Flare	GPT Image 2.5 Flare	GPT Image 2.5 Flare (max)
3	GPT Image 2 (medium)	GPT Image 2 (medium)	GPT Image 2 (high)
4	MAI-Image-2.6 (Microsoft)	Grok Imagine Image 2.0	Grok Imagine Image 2.0
5	Reve 2.1	MAI-Image-2.6	MAI-Image-2.6
6	Grok Imagine Image 2.0	Muse Image (Meta)	Nano Banana 2
7	Muse Image (Meta)	MAI-Image-2.5	Muse Image
8	Reve 2.0	Seedream 5.0 Pro	GPT Image 1.5 (high)
9	Nano Banana 2	Nano Banana Pro (2K)	MAI-Image-2.5
10	Seedream 5.0 Pro	Grok Imagine Image Quality	Nano Banana Pro

Arena boards dated September 24 and 21, 2026; Artificial Analysis board as shown in September 2026.

Leaderboards reward images people like at a glance, from prompts that are not your product. They cannot tell you whether your label came out right. Use them to choose which two or three models to test.

Which model for which product job

Job	Start with	Why
Keep a packshot the same while the scene changes	GPT Image 2.5 Sunburst, then Nano Banana Pro	OpenAI built Sunburst for editing precision, and it leads Arena’s edit board. Nano Banana Pro takes up to 6 object references.
Labels and short text on packs	GPT Image 2.5 or Nano Banana Pro	Google lists advanced text rendering for Nano Banana; OpenAI says text is much improved but placement can slip. Check every letter at 100%.
Many references in one shot: product, model and style	Nano Banana 2	Up to 10 object, 4 character and 3 style references per request
Cheap drafts and many variations	GPT Image 2 at low quality, Muse Image, or Nano Banana 2 Lite	$0.006, $0.01 and $0.034 an image
4K for print and billboards	Nano Banana 2 or Nano Banana Pro at 4K	$0.151 and $0.24 an image; GPT Image allows up to 3,840 pixels on the long edge
A transparent background	GPT Image 2 or 2.5	Transparent PNG and WebP output
Mood and styled campaign images	Midjourney V8.2 or GPT Image 2.5	Midjourney’s V8.2 update focused on aesthetics

For large prints, check the size a billboard actually needs in billboard resolution for AI images before you pay for 4K on every render.

What every model still gets wrong

No model copies your product exactly. In OpenRouter’s September 2026 test, four models were told to keep a coffee bag and its printed label exactly and change only the background. All four kept the words and the typeface; all four changed the bag’s shape, proportions or lighting to some degree, and none reproduced it pixel for pixel. ByteDance says Seedream 5.0 Pro still has room to improve on fine text and pixel-level editing, and OpenAI says its models can lose consistency for recurring brand elements across generations.

So plan for it. Give each shot only the references it needs, say what must not change, check at full size, and when a label must be exact, put the real packshot or label into the image afterward. See why AI changes your product and AI photos for packaged goods.

How to test models on your own product

Pick three models: one from the top of the leaderboards, one cheap one, and one you already pay for.
Pick three shots you need: a packshot on white, a lifestyle scene, and a shot with a person or with text.
Use the same reference photos for every model. Write one prompt per shot and fit its length to the model: the prompts behind our agency’s kept photos ran 140 to 200 words on Nano Banana and about 300 on GPT Image.
Render each shot 10 times on each model.
Mark the keepers: the product matches, nothing is warped, and the image works without retouching. Count them.
Divide what you spent by the number of keepers, and compare models on that cost per usable photo. Our own campaigns averaged about 8 renders per keeper; see how many AI renders it takes to get one usable photo.

Model test sheet

Product: ____________   Date: ____________
References used (the same for every model): front / back / label / texture

Shot               | Model | Renders | Keepers | Cost per render | Cost per keeper | Notes
Packshot on white  |       | 10      |         |                 |                 |
Lifestyle scene    |       | 10      |         |                 |                 |
Person or text     |       | 10      |         |                 |                 |

A keeper: label text matches, logo shape matches, color matches,
proportions match, no extra parts, usable without retouching.

The same model costs different amounts in different tools

Apps resell these models inside their own credits and add their own features, so one Nano Banana Pro image costs a different amount depending on where you run it.

Where you run it	One Nano Banana Pro image	What else you get
Google’s Gemini API	$0.134 (1K or 2K)	The model only
OpenRouter	$0.134 (measured at 1024 × 1024)	One account for many makers’ models
Higgsfield	2 credits, about $0.07 to $0.14 by plan	Templates, video models, a web studio
Flora	$0.18 (1K)	A node canvas and team seats
Uwear	5 credits, $0.50	Fashion tools; an optional check against the product photo costs 1 credit more

From each vendor’s price page or published test, September 2026.

OpenRouter passes the makers’ prices through without markup and adds its fee when you buy credit; see what OpenRouter is.

Questions people also ask

Is Nano Banana the same as Gemini?: Nano Banana is Google’s name for the image models in its Gemini family. Nano Banana 2 is Gemini 3.1 Flash Image, Nano Banana 2 Lite is Gemini 3.1 Flash Lite Image, and Nano Banana Pro is Gemini 3 Pro Image. The Gemini app uses Nano Banana 2 for everyone and offers Nano Banana Pro on paid plans.
What happened to Imagen?: Google shut down Imagen 4 on the Gemini API on August 17, 2026 and recommends Nano Banana 2 (gemini-3.1-flash-image) in its place.
Which model is cheapest for product photos?: Per render, GPT Image 2 at low quality ($0.006 in OpenRouter’s test), Meta’s Muse Image ($0.01 on OpenRouter) and Nano Banana 2 Lite ($0.034) are the cheapest well-known options in September 2026. The cheapest per usable photo depends on how many renders you throw away, so count keepers before you switch.
Why does GPT Image cost so little in one test and so much in another?: OpenAI prices GPT Image by tokens, and the quality setting changes how many tokens an image uses. OpenRouter found that leaving quality unset billed the same as low, $0.006 for a square image, while high cost $0.21. Set the quality on purpose whenever you compare prices.

Where Overs fits

Overs lets you choose the model that makes the final photos: Nano Banana 2 Lite, Nano Banana 2 (the default), Nano Banana Pro, GPT Image 2.5 Flare or Sunburst, or GPT Image 2. You can draft on a cheap model and use "Finish the keepers" to remake the photos you approved on a better one. Model costs go on your own key from OpenRouter, a service that bills many AI models through one account; the Overs plan is free for 40 photos a month.

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Start free How it works

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.