Nano Banana 2.1: Is It Free, Price per Image, vs Nano Banana 2 and Pro, and the API

Nano Banana 2.1 is Google's newest image generation and editing model, released on October 6, 2026. It is the update to Nano Banana 2, it runs on Gemini 3.6 Flash, and on the Gemini API it costs $0.0336 for a 1K image and $0.0504 for a 2K image, half of Nano Banana 2's price. You can use it for free in the Gemini app and in AI Mode in Search, within each plan's limits, but the API has no free tier. Here is the part almost nobody is saying: in Google's own blind comparisons, this budget model was preferred over Nano Banana Pro in every category Google published, from infographics to keeping a character's face the same, while costing a quarter of Pro's price per image. The catch is on the other side of the bill: its input price tripled, so editing with lots of reference photos saves much less than half.

Jake runs a phone repair shop, and every few weeks he needs a poster for the window: this month it was "Screen Repair Saturday, any screen $49, done while you wait." He had been making them in the Gemini app with Nano Banana 2, and the price always came out slightly wrong, a "$94" here, a missing dollar sign there, so he ended up fixing the text in a photo editor. When he read that Nano Banana 2.1 renders text better, he asked Ethan, who builds things on Google Cloud and AWS for a few local businesses, whether it was worth changing anything. Ethan had spent the morning moving a furniture client's product-photo pipeline to the new model. "For your poster, nothing changes except the result. For my client, the price per image halved and the price per reference photo tripled, and I had to rewrite two lines of code." This page covers both sides: what Nano Banana 2.1 is, whether it is free, what each image really costs, how it compares with Nano Banana 2, Lite and Pro, working API code, what changed for developers, and what to do about the older models.

If "Nano Banana" still sounds like a joke name, that is fair. It began as the nickname for Google's first Gemini image model, people kept using it, and Google adopted it officially. Today it is the family name for every image model in Gemini, and there are five of them, which is why searches for "Nano Banana 2", "Nano Banana Pro" and "Nano Banana AI" all land in slightly different places. The table in the next section untangles them in one look.

What is Nano Banana 2.1? The family, in one table

Nano Banana 2.1 is a "native" image model, which means it is a Gemini language model that can also draw. You describe what you want in plain words, or hand it a photo and describe the change, and it returns an image, often with a short text reply. Because it is built on a language model, it understands long, specific instructions, can look things up with Google Search before drawing, and can keep editing the same picture over several turns of conversation.

Here is the whole Nano Banana family on the Gemini API, newest first:

What Nano Banana 2.1 adds over Nano Banana 2, in Google's own list: better visual quality and realism at 1K, 2K and 4K; fixed tiling artifacts on very wide and very tall images (the 1:4, 4:1, 1:8 and 8:1 shapes) at 2K and 4K; better text rendering and infographic layout; up to 14 reference images in one request, keeping up to 4 characters consistent and up to 10 objects faithful; grounding with Google Web Search and Image Search; and three thinking levels, minimal, medium and high.

The specifications that matter in practice: it takes text, images, video and PDFs as input; it returns images and text; the API accepts up to 131,072 input tokens and up to 32,768 output tokens per request; it supports 15 aspect ratios, from 1:1 and 16:9 to 21:9 and 1:8; and every output carries Google's SynthID watermark. It does not take audio, does not call functions, and does not return structured JSON.

Its brain is Gemini 3.6 Flash, with a knowledge cutoff of March 2026. That matters for prompts about recent things: if you ask for a poster about something that happened after March, turn on Search grounding so the model can look it up instead of guessing.

Is Nano Banana free? Where you can use 2.1 without paying

This is the most common question, so here is the plain answer. Nano Banana is free in Google's consumer apps and paid on the developer API.

• The Gemini app (web, Android, iPhone): image generation with Nano Banana is included on every plan, including the free one. The free plan has "standard limits". Google AI Plus is 2 times higher, Google AI Pro 4 times, and Google AI Ultra 5 or 20 times higher than Pro, depending on the Ultra tier. Limits are based on compute, so a 4K image or a heavy edit uses more of your allowance than a simple square picture. You can see where you stand under Settings, then Usage Limits, on gemini.google.com.
• "Redo with Nano Banana Pro" in the Gemini app needs a paid Google AI plan. The free plan does not include it.
• AI Mode in Google Search also offers Nano Banana 2.1 for free, within limits.
• Other Google products: Google lists Google AI Studio, Google Flow, Google Stitch and Google Ads among the places Nano Banana 2.1 is rolling out.
• The Gemini API: no free tier for Nano Banana 2.1, or for any current Nano Banana model. You need a billing-enabled API key, and you pay per image.

One honest timing note. Google lists the Gemini app as a home for Nano Banana 2.1, but the Gemini app's own limits page still named Nano Banana 2 on October 7. Rollouts in the app usually reach accounts gradually, so if your results look the same as last week, give it a few days; nothing on your side needs changing.

A second date to keep in mind: from October 9, 2026, Google moves the free Gemini app to its Flash-Lite text model, while paid plans keep Flash and Pro. Google's feature table still lists image generation with Nano Banana on the free plan, but if you rely on the free app for images, check your results after the 9th. Our Gemini free tier guide covers that change in full.

Jake's poster went through the free Gemini app, the same way as always: "A bright shop-window poster for a phone repair shop. Big headline: SCREEN REPAIR SATURDAY. Below it: Any screen, $49, done while you wait. Clean modern style, blue and white." Ethan's only tip was to keep the text short and put it in quotation marks. The price came out right on the first try.

Before you upload photos of people, especially children, to any AI app, check what the app keeps. In Gemini that lives in your Gemini Apps Activity settings. Our piece on why viral AI photo trends deserve a second thought explains what to look for in any app.

Nano Banana 2.1 pricing: the price per image, and the input catch

On the Gemini API, image output is billed in tokens, and Google publishes what that works out to per picture. A 1K image (1024 by 1024 pixels) uses 1,120 output tokens, a 2K image 1,680 tokens, and a 4K image 3,780 tokens. At $30 per million image tokens, that gives the prices below:

Gemini API, paid tier, US dollars, as of October 7, 2026.

Read the table twice, because the headline "half price" is true in one place and misleading in two others.

Where it is true: at 1K and 2K, Nano Banana 2.1 costs almost exactly half of Nano Banana 2, and a quarter of Nano Banana Pro at 1K.

Where it is not: at 4K, 2.1 uses 3,780 tokens per image against Nano Banana 2's 2,520, so a 4K picture is only about 25% cheaper, $0.113 instead of $0.151. And 2.1 no longer offers the small 512-pixel size that Nano Banana 2 had at $0.045, so thumbnail pipelines move up to 1K.

The input catch: input now costs $1.50 per million tokens instead of $0.50, and text and thinking output $7.50 instead of $3.00. Every reference image you send costs 1,120 input tokens, about $0.0017 each. That is nothing for a one-photo edit. For a request with 14 reference images, input alone is about $0.024, which is most of a second picture. Two more details push the same way: thinking is on by default at medium (Nano Banana 2 defaulted to minimal), and thinking text is billed at the output text rate. The interim "thought images" the model draws while reasoning are not charged.

Batch is half price. If the images do not need to come back within seconds, the Batch API halves everything: $0.0168 per 1K image, $0.0252 per 2K and $0.0567 per 4K, with input at $0.75. Overnight catalog jobs belong there. Flex and Priority processing are not offered for Nano Banana 2.1.

Search grounding is billed separately: 5,000 grounded requests a month are free across all Gemini 3.x models, then $14 per 1,000. Text and images the search returns are not charged as input tokens.

Interesting detail for the cost-conscious: at 1K, Nano Banana 2.1 now costs exactly what Nano Banana 2 Lite costs. Lite is still the cheaper choice when inputs are heavy and speed matters most, because its input price is $0.25, but on a simple text-to-image call the "Lite" discount has disappeared.

What 1,000 images really cost: two worked examples

Per-image prices are easy to quote and easy to misjudge, so here are the two jobs from the opening story, worked through on the Gemini API. Thinking text is left out because its length varies by prompt; it adds a little to every line.

Example 1: a product catalog at 2K. Ethan's furniture client needs 1,000 lifestyle photos, each made from 2 product photos plus a 200-token prompt. Input per request is 2 × 1,120 + 200 = 2,440 tokens.

For a catalog like this, 2.1 really is about half of Nano Banana 2, and the Batch API halves it again. Moving to Nano Banana 2.1 Batch took the client's monthly bill from roughly $100 to roughly $27 for the same 1,000 pictures.

Example 2: a heavy edit at 1K. A designer combines 14 reference images into one 1K scene. Input is 14 × 1,120 + 200 = 15,880 tokens, which costs about $0.024 on Nano Banana 2.1. Add the $0.0336 picture and the request costs about $0.057. That is still cheaper than Nano Banana 2's output price alone at 1K, $0.067, but the saving is closer to a fifth than a half. The more reference images you send, the less the headline discount applies.

If you need to estimate before you commit, run twenty real requests and read the usage numbers in each response, then multiply. The model's thinking length varies with the prompt, and that is the one number a table cannot give you.

Nano Banana 2.1 vs Nano Banana 2 vs Pro: which one to use

Google's model card for Nano Banana 2.1 includes side-by-side human preference scores, where raters pick the better of two images without knowing which model made them. Higher is better. These are Google's own tests, so read them as the vendor's evidence, not an independent review:

Two things stand out. Nano Banana 2.1 with thinking beat every other model in every row, including Pro. And the jump from turning thinking on is real, especially for factuality in infographics, where the share of correct results went from 0.328 to 0.521. That is the case for keeping the default medium thinking, or using high, on anything with facts or numbers in it.

So why does Pro still exist? Google positions it as the premium choice for the most complex work: the deepest world knowledge, advanced localization, accurate brand consistency and precise creative control. Pro is also the only model that accepts up to 3 dedicated style reference images, and it keeps up to 5 characters consistent against 2.1's 4. If you produce brand assets where one logo color being off is a problem, test both before switching.

A simple way to choose:

• Nano Banana 2.1 for almost everything: posters, product shots, infographics, edits, anything with text in it, 1K to 4K.
• Nano Banana 2 Lite for speed and volume at 1K, such as stickers, quick background swaps or live previews in an app, where a sub-two-second answer matters more than the best picture.
• Nano Banana Pro for brand-critical work that needs style references or more than four consistent characters, and where you have tested that it beats 2.1 on your own images.
• Nano Banana 2 only while you migrate. Google recommends 2.1 for all new projects.

How to make your first Nano Banana 2.1 image with the API, step by step

If you have never used the Gemini API, this is the whole path from nothing to a saved picture. It takes about ten minutes, most of it spent on billing.

1. Sign in to Google AI Studio at aistudio.google.com with the Google account you want to bill.
2. Create an API key from the API keys page. AI Studio creates or picks a Google Cloud project to hold it.
3. Turn on billing for that project. This step is not optional for image generation: the Nano Banana models have no free API tier, so a key without billing cannot make images. Set a budget alert while you are there.
4. Try a prompt in AI Studio first. Choose gemini-nano-banana-2.1 as the model, type a description, and look at the result before writing any code. It is the fastest way to learn what the model does well.
5. Install the SDK and set the key on your computer, as shown below, then run the first example.
6. Check what you were charged. Each response reports its token usage. Multiply a few real requests out to your expected monthly volume before you put the model into anything that runs on its own.

Keep the key out of your code and out of anything you share. If a key leaks, delete it in AI Studio and create a new one; a leaked billing-enabled key can run up real charges on images someone else is generating.

How to use Nano Banana 2.1 with the Gemini API: working code

With a billing-enabled key from the steps above, install Google's Gen AI SDK for Python and set the key:

pip install --upgrade google-genai pillow
export GEMINI_API_KEY="your-key-here"

The simplest request: describe the picture, save the result.

import base64
from google import genai

client = genai.Client()   # reads GEMINI_API_KEY

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input='A shop-window poster for a phone repair shop. Headline: "SCREEN REPAIR SATURDAY". '
          'Below it: "Any screen, $49, done while you wait". Clean modern style, blue and white.',
)

with open("poster.png", "wb") as f:
    f.write(base64.b64decode(interaction.output_image.data))

To choose the shape and size, add response_format. The size must be written with a capital K, as 1K, 2K or 4K; a lowercase 2k is rejected.

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="A wide banner of a cozy phone repair workbench at golden hour, room on the left for text.",
    response_format={
        "type": "image",
        "mime_type": "image/jpeg",
        "aspect_ratio": "16:9",
        "image_size": "2K",
    },
)

Editing a photo works the same way, with the image passed in alongside the instruction:

with open("shop_front.jpg", "rb") as f:
    photo = base64.b64encode(f.read()).decode("utf-8")

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input=[
        {"type": "text", "text": "Replace the old sign above the door with a clean blue sign "
                                 "that says JAKE'S PHONE REPAIR. Keep everything else the same."},
        {"type": "image", "data": photo, "mime_type": "image/jpeg"},
    ],
)

Keep editing in conversation by passing the previous interaction's ID. This is Google's recommended way to refine an image, because the model remembers what it drew:

second = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="Same image, but make the sign text white and add a small phone icon before the name.",
    previous_interaction_id=interaction.id,
)

Set the thinking level when you want more care, or more speed. The default for 2.1 is medium; high helps with infographics, diagrams and anything factual, and minimal is fastest:

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="An infographic showing how a phone screen is replaced, in five labeled steps.",
    generation_config={"thinking_level": "high"},
)

Ground it in real information by adding the Google Search tool, so the model can check facts before drawing, which is useful for anything about recent events, weather or places:

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="A simple poster showing this weekend's weather forecast for Austin, Texas.",
    tools=[{"type": "google_search"}],
)

By default the model returns text and an image together. To get only the image, ask for an image format in response_format as in the second example. Video input works too: you can pass a video, including a YouTube link, and ask for a poster or thumbnail that captures it.

What changed for developers moving from Nano Banana 2

For most code, migration is one model name: gemini-3.1-flash-image becomes gemini-nano-banana-2.1. These are the differences that can still break a working pipeline or a budget:

1. No 512-pixel size. Nano Banana 2 offered 0.5K images; 2.1 starts at 1K. A thumbnail pipeline that asks for 0.5K needs a new size, and its cost per image changes accordingly.
2. Thinking now defaults to medium. Nano Banana 2 defaulted to minimal. If latency or text-output cost matters more than quality for a route, set minimal explicitly.
3. Input is three times the price. Workloads with many reference images save less than the headline suggests. Re-run your cost estimate with 1,120 tokens per input image.
4. No Flex or Priority. Standard and Batch are the only ways to buy it on the Gemini API.
5. Sampling settings are refused. On Google Cloud, setting temperature, topP, topK, seed or logprobs for Nano Banana 2.1 returns an API error. Remove them from shared request templates.
6. Grounding and real people. Search grounding on 2.1 does not use real-world images of people from web search.

On Google Cloud (now called Gemini Enterprise Agent Platform, formerly Vertex AI), the model ID is the same, gemini-nano-banana-2.1. It is served from the global endpoint only, supports Provisioned Throughput and Batch, and adds a few enterprise extras: implicit context caching, token counting, Content Credentials (C2PA), virtual try-on and person generation. Per request you can send up to 14 images, each up to 7 MB inline or 30 MB from Cloud Storage. Data residency, CMEK and VPC Service Controls are listed among its security controls, though the global-only endpoint is worth discussing with whoever owns your data-location promises.

On Firebase AI Logic, for mobile and web apps, the Nano Banana models require the pay-as-you-go Blaze plan, whichever Gemini API provider you choose.

Ethan's migration for the furniture client came down to three edits: the model name, removing a leftover temperature setting from a shared template, and moving the overnight job to Batch. The first test images were sharper at the edges of wide room shots, which is the tiling fix in practice.

Nano Banana 2.1 not working? Common mistakes and the fixes

Most "it doesn't work" moments with Nano Banana come from a short list of causes. Work through them in order:

1. No billing on the API key. The Gemini API has no free tier for Nano Banana. Turn on billing for the key's project; the free route is the Gemini app, not the API.
2. An old model name. Retired names stop working entirely, and a request using one returns a 404. The common ones are gemini-3.1-flash-image-preview, gemini-3-pro-image-preview, gemini-2.0-flash-preview-image-generation and the Imagen 4 models. Switch to gemini-nano-banana-2.1.
3. A lowercase size. image_size must be 1K, 2K or 4K with a capital K. 2k is rejected.
4. Asking 2.1 for a 512-pixel image. The 0.5K size belonged to Nano Banana 2. Use 1K.
5. Sampling settings in the request. temperature, topP, topK, seed and logprobs are not supported, and Google Cloud returns an API error when any is set. Remove them.
6. Too many or too large reference images. The limit is 14 images per request. On Google Cloud, inline images can be up to 7 MB each and Cloud Storage files up to 30 MB.
7. Using Lite for what only 2.1 does. Nano Banana 2 Lite makes 1K images only and does not support Search grounding. Use 2.1 for 2K, 4K or grounded images.
8. Grounded images of real people. Search grounding on 2.1 does not use real-world images of people from web search, so a "picture of this celebrity at today's event" will not be built from news photos.
9. Fewer images than requested. The model does not always return the exact number of images asked for. Make one request per image when the count matters.
10. Blurry or misspelled small text. Use 2K, shorten the text, put it in quotation marks, and set thinking to high. Write the text first, then ask for the image with it.
11. The Gemini app still looks like the old model. The app is getting 2.1 gradually. Nothing in your settings switches it on early.

If an edit keeps changing things you wanted kept, say so explicitly: "Keep everything else in the image exactly the same, preserving the original style, lighting and composition." That sentence, straight from Google's own templates, fixes more edits than any setting.

Is Nano Banana 2 being shut down? The real dates

If you have read that Nano Banana 2 shuts down on October 29, 2026, slow down before you rush a migration. Several sites are reporting that date, but on October 7 Google's own pages said something different:

• Gemini API changelog, October 6: "The gemini-3.1-flash-image model is deprecated (no shutdown date announced). Migrate to gemini-nano-banana-2.1."
• Gemini API deprecations table: gemini-3.1-flash-image, no shutdown date announced, recommended replacement gemini-nano-banana-2.1.
• Google Cloud model page (updated October 6): retirement date "May 28, 2027 or later".
• Firebase AI Logic: shutdown "no earlier than" May 28, 2027.

"Deprecated" means Google wants you to move and will eventually set a date; it does not mean the model stops working this month. The sensible plan is to test 2.1 now, switch when your results are good, and watch the deprecations page for a real date. If Google does announce October 29, you will have already moved.

The older models have firmer dates, and some have already gone:

The original Nano Banana deserves a warning of its own. Google's deprecations table now lists March 15, 2027, but its pricing page still warns of an October 2, 2026 shutdown, and on Google Cloud the Gemini 2.5 models are being shut down for all projects in October 2026. When two official pages disagree, plan for the earlier date. If anything you run still calls gemini-2.5-flash-image, move it to Nano Banana 2 Lite or 2.1 this week.

Nano Banana prompts that work: Google's own templates

Google's image guide includes fill-in-the-blank templates for the jobs people do most. They are worth copying, because they show what the model listens to: the kind of shot, the light, the camera, the purpose and, for edits, what must not change. In short form:

Jake's version of the product-photo template, for a refurbished phone listing: "A high-resolution, studio-lit product photograph of a refurbished blue smartphone on a white marble counter. Soft three-point lighting to show the glass without glare. Three-quarter angle to show the camera bump. Sharp focus on the screen. Aspect ratio 4:5." He used the result on his shop's social page the same afternoon.

Two habits make the templates work better. First, say what the image is for: "a logo for a minimalist skincare brand" gets a better result than "a logo". Second, refine in the same conversation with small requests, such as "the same, but warmer light", instead of rewriting the prompt from scratch, because the model keeps what it already got right.

What Nano Banana 2.1 still gets wrong

Google's own model card is unusually frank about the weak spots, and knowing them saves a lot of regenerating:

• Small text is still blurry, especially at 1K. Headlines and prices come out well; fine print, long paragraphs and full pages of text do not. Keep text short, put it in quotation marks in the prompt, and use 2K for posters with smaller lines.
• Character consistency is good, not perfect. Faces can drift between input and output, so check every image before you publish a series.
• Doodle and mask edits are partly followed. Drawing on a photo to mark the change works better than before but not every time.
• Poses can stick. When editing, the model sometimes keeps a person's original pose when you asked for a new one.
• Left and right get confused. Describe positions with landmarks ("next to the door") rather than "on the left".
• World knowledge and 3D reasoning are limited. Diagrams of how things work can look convincing and be wrong. Turn on thinking and Search grounding for anything factual, and check the result.
• The number of images is not guaranteed. Ask for four variations and you may get three.

Google's own prompting advice fits on a sticky note: be specific, say what the image is for, describe the camera ("wide-angle shot", "low angle"), describe what you want instead of what you do not want ("an empty street" rather than "no cars"), and refine in small steps in the same conversation. For text in images, Google suggests writing the text first and then asking for the image with it.

SynthID: every Nano Banana image is labeled

Every image from every Nano Banana model includes a SynthID watermark, an invisible signal woven into the pixels that Google's tools can detect even after cropping or light editing. You cannot switch it off. On Google Cloud, Nano Banana 2.1 also supports Content Credentials (C2PA), the open standard that records how an image was made.

For a small business like Jake's, that changes nothing; a poster is a poster. For anyone tempted to pass an AI image off as a real photo, it is a reason not to. And for readers, it is reassuring: the Gemini app can check whether an image was made with Google AI, which helps when a "photo" in a group chat looks too perfect.

Can you run Nano Banana locally?

No. Every Nano Banana model is closed. Google has not released the weights, so there is no download and no Ollama tag for the real model. Tools that offer Nano Banana inside other apps are calling Google's API behind the scenes, and anything offering a "Nano Banana download" is a different model wearing the name. It is not offered on Amazon Bedrock; outside Google's own apps, you reach it through the Gemini API or Google Cloud, or through services that resell that access.

If you want image generation on your own computer, for privacy, for unlimited use, or just to learn, open-weight image models are the route. Our guide to installing Qwen-Image-2.1 on Windows and Kali walks through one through ComfyUI, the node-based tool most local image work runs on, and our YuE2 ComfyUI guide shows the ComfyUI install itself. Expect to need a graphics card with plenty of memory; the laptop tiers guide explains what is realistic.

Which Nano Banana route should you use?

• You want a few images for yourself or your shop: the free Gemini app. Upgrade to a Google AI plan only if you hit the daily limit often or want "Redo with Nano Banana Pro".
• You are building an app: the Gemini API with gemini-nano-banana-2.1, thinking at the default, and the Batch API for anything that can wait.
• You need speed above all, at 1K: Nano Banana 2 Lite.
• Your company runs on Google Cloud: the same model on Gemini Enterprise Agent Platform, with Provisioned Throughput, C2PA and enterprise controls, on the global endpoint.
• You still call Nano Banana 2, Imagen or the original Nano Banana: test 2.1 now. The original Nano Banana is the urgent one.
• You want it on your own hardware: not possible with Nano Banana; use an open model in ComfyUI instead.

New model names every few months can make it feel as if you are always one version behind. You are not. For most people, Nano Banana 2.1 simply means better pictures from the same Gemini app, and for developers it is a one-line change with two settings to check. Jake's window poster now says $49 the first time, and it has been in the window all week. Ethan's furniture client got sharper room shots for about a quarter of last month's bill. Try one real image of your own on the new model; the result will tell you more than any chart.

Revision note. Written October 7, 2026, the day after Nano Banana 2.1 became generally available. If a headline sent you racing to migrate before October 29, breathe out; you have time to test properly.