Nano Banana Pricing & Review: Google's Gemini Image Models

"Nano Banana" is Google DeepMind's codename for its Gemini image generation and editing family — not one model but a tier: Nano Banana Pro (gemini-3-pro-image), Nano Banana 2 (gemini-3.1-flash-image), a fast Lite, and the original Nano Banana (Gemini 2.5 Flash Image). Pricing is token-based but effectively per image: Google prices each image by its resolution, so a 4K render costs more than a 1K one. This guide covers what each tier is, exactly what an image costs (dated and unit-labelled), the vendor-reported quality features, and how to call Nano Banana through an OpenAI-compatible endpoint. Prices are as of July 2026 — generative-media pricing moves, so treat the source links as canonical.

What Nano Banana is

"Nano Banana" is Google DeepMind's codename for its Gemini image generation and editing models — not a single product but a family. The name first attached to the original Nano Banana, which is Gemini 2.5 Flash Image, announced in August 2025. Since then Google has shipped higher and lower tiers under the same banana branding, and the name has stuck as the popular label for "Gemini's image model."

Practically, when someone says "Nano Banana" in 2026 they usually mean one of three current tiers — Pro, 2, or Lite — plus the original. They share the Gemini backbone (so the model understands your prompt with real world knowledge and can edit as well as generate), and they all differ mainly in resolution, latency, and price. If you want the wider text-to-image landscape rather than just Google's family, see our best text-to-image models of 2026 roundup.

The Nano Banana model tiers

There are four Nano Banana models worth knowing, from the flagship Pro down to the original. Here is what each name maps to:

• Nano Banana Pro — gemini-3-pro-image (also gemini-3-pro-image-preview). Built on Gemini 3 Pro; announced Nov 20, 2025 as Google DeepMind's new state-of-the-art image generation and editing model. Supports 2K and 4K output. This is the "maximum control" tier.
• Nano Banana 2 — gemini-3.1-flash-image. Google describes it as "the generalist workhorse," delivering high quality at lower latency and, per Google, "the best balance of performance and cost." The everyday default.
• Nano Banana 2 Lite — gemini-3.1-flash-lite-image. Google's fastest, most cost-efficient Gemini image model: text-to-image in about 4 seconds while retaining prompt adherence, character consistency, and legible in-image text (vendor-reported).
• Original Nano Banana — Gemini 2.5 Flash Image (Aug 2025). Still available and cheap at $0.039 per image; superseded on quality by the Gemini 3 generation above.

Because the tiers share one API surface, choosing between them is mostly a price/resolution decision rather than a re-integration — you change the model string, not your code.

How much Nano Banana costs

Nano Banana is billed per image, priced by resolution — Google publishes it as token-based, but each image maps to a fixed output-token count, so you effectively pay a known price per picture. Your prompt text and any reference images are billed separately as input tokens. All figures below are from Google's official Gemini API pricing page, verified July 2026; a 50%-off Batch tier exists for non-real-time work.

Two things to keep straight. First, the unit is per image, not per prompt — asking for four variations bills four images. Second, resolution is the price lever: a Nano Banana Pro 4K render (~$0.24) costs nearly double its 1K render (~$0.134), so generate previews small and only upscale the keepers. For how this compares against OpenAI GPT Image and other providers on a cost basis, see the best AI image API guide and the broader cheapest-API cost analysis.

Nano Banana Pro vs 2 vs Lite: the decision table

Pick Pro for control and 4K, Nano Banana 2 for the everyday balance, Lite for speed and volume. This synthesised table lines up the tiers on the axes that actually drive the choice:

Latency and "positioned as" rows are vendor-reported (Google DeepMind); resolution and price are from the official pricing and announcement pages linked below. We do not publish benchmark scores we have not run.

Image quality and features

Google positions Nano Banana Pro as offering "unprecedented control, improved text rendering and enhanced world knowledge" over the original — those are the vendor's words, and they map to concrete features. From Google DeepMind's Nano Banana Pro announcement (Nov 20, 2025), the Pro tier can:

• Render legible in-image text in multiple languages — the classic weak spot for image models, and the headline improvement here.
• Edit locally — select, refine, or transform any part of an image rather than regenerating the whole thing.
• Adjust camera angle, focus, and color grading after the fact.
• Blend up to 14 images into one composition.
• Keep up to 5 people consistent across a set of images — useful for character or brand continuity.

These are capability claims from the vendor, not independent benchmarks. Because all four tiers sit on the Gemini backbone, they inherit its world knowledge — so prompts that reference real objects, brands, or scenes tend to come out more coherent than with a pure-diffusion model. If you are choosing an image model on evidence rather than marketing, weigh these features against the field in our text-to-image comparison.

How to call Nano Banana (with code)

Through DataLLM Lab you call Nano Banana with the same OpenAI-compatible client you use for chat — name the model, read back the image. There is no separate Google Cloud project, no Vertex SDK, and one API key covers both your text and image models. The base URL is https://www.datallmlab.com/v1:

from openai import OpenAI

client = OpenAI(
    base_url="https://www.datallmlab.com/v1",
    api_key="YOUR_DATALLMLAB_KEY",
)

# Nano Banana Pro = gemini-3-pro-image; Nano Banana 2 = gemini-3.1-flash-image
resp = client.images.generate(
    model="gemini-3-pro-image",
    prompt="A product hero shot of a matte-black water bottle on wet slate, "
           "studio lighting, the label reads 'HYDRA' in crisp sans-serif",
    size="2048x2048",          # 2K; use 4096x4096 for 4K (higher per-image cost)
    n=1,                       # each image bills separately — preview small first
)

image_url = resp.data[0].url
print(image_url)

Swap model to gemini-3.1-flash-image for the cheaper Nano Banana 2 tier — nothing else changes. Because the endpoint is OpenAI-compatible, any SDK, cURL call, or framework that already targets the OpenAI images API works unmodified; just point base_url at the gateway. Grab a key on the register page.

SynthID watermark and honest caveats

Every Nano Banana image carries an invisible SynthID watermark, and the per-1M-token prices you see displayed do not map cleanly to a per-image model — know both before you build. Two honesty notes:

• SynthID. All images generated or edited with Gemini image models include an invisible SynthID digital watermark identifying them as AI-generated or AI-edited (per Google's original Nano Banana announcement). It is imperceptible in the picture but detectable by SynthID tooling — relevant if provenance or disclosure matters to you.
• Per-token vs per-image. Google prices images by a fixed output-token count per resolution (Pro 1K = 1120 tokens, 4K = 2000 tokens), which implies a headline rate around $120/1M output tokens for Pro. A gateway catalog that lists a model at a nominal per-1M output rate (DataLLM Lab shows gemini-3-pro-image at $2 in / $12 out per 1M, and gemini-3.1-flash-image at $0.50 / $3) is a display convention — image models are practically billed per image, so use the per-image numbers in the pricing table above when you budget.

Calling Nano Banana on DataLLM Lab

The gateway's value here is one integration, not a cheaper banana. You call gemini-3-pro-image and gemini-3.1-flash-image through the same https://www.datallmlab.com/v1 endpoint and key that serves your chat models, so an app that already talks to Claude, GPT, or Gemini can add image generation without a second vendor account, a second billing relationship, or a second SDK. If you want the mechanics of why a single OpenAI-compatible surface simplifies a multi-model stack, read what an LLM gateway is; for the full serving catalog see the best LLM API of 2026. Note that text-to-video and world models are a separate landscape — DataLLM Lab serves chat and image today, not video — so if you are exploring generative video, our text-to-video guide covers that field independently.

FAQ