Google launches Nano Banana 2.1, a better image model at half the price

Nano Banana 2.1 is Google's new Flash-tier image model. "Flash" is Google's label for its fast, low-cost models, as opposed to the bigger and slower "Pro" ones.

According to Google's model card, it's built on Gemini 3.6 Flash. It takes text and images as input and returns images and text. Its API name is gemini-nano-banana-2.1.

It replaces Nano Banana 2, which Google launched in February 2026. That earlier model was officially called Gemini 3.1 Flash Image.

Google says the new version improves on previous ones "across the board," The Decoder reports. The company calls it "the more efficient counterpart" to Nano Banana Pro, its top image model.

Google's developer documentation lists these upgrades:

• Visual quality: more natural-looking images and better visual design.
• Text rendering: words inside images, like signs, labels and posters, come out cleaner.
• Multi-turn consistency: a character stays the same person across several edits in one conversation.
• Mask-based editing: you mark one area of a picture, and only that area changes.
• Infographics and panoramas: charts, diagrams and wide scenes, The Decoder notes.

The reference image limit is generous. You can feed it up to 14 images in one request. Google says it keeps up to 4 characters and up to 10 objects consistent with those references.

The model can also look things up. It connects to Google Search, including Google Image Search, so it can check what a real product, landmark or object looks like before drawing it.

In real life Say you run a bakery. You upload photos of your three best cakes and your logo, then ask for a fall menu poster. The model can keep the cakes looking like your cakes and, with better text rendering, spell "pumpkin" correctly.

This is where the news gets concrete. Here are Google's API prices per image, standard tier:

By the numbers: every size drops by roughly 50%. Batch requests, which you send in bulk and get back later, cost half again: $0.0168 for a 1K image.

For comparison, Nano Banana Pro costs $0.134 per 1K or 2K image and $0.24 per 4K image. So a 1K image from 2.1 costs about a quarter of what Pro charges.

The catch: the per-image price fell, but Google's pricing page shows the input price rose from $0.50 to $1.50 per million tokens. Tokens are the small chunks of text and image a model reads. For most image jobs the output image is the big cost, so the total still drops. There is also no free tier in the API.

A few, and the names can get confusing. Here's the family, with each model's official name, as The Decoder lists it:

• Nano Banana: Gemini 2.5 Flash Image, the original.
• Nano Banana Pro: Gemini 3 Pro Image, the top-quality model.
• Nano Banana 2: Gemini 3.1 Flash Image, launched in February 2026.
• Nano Banana 2 Lite: Gemini 3.1 Flash Lite Image, the smallest one.
• Nano Banana 2.1: the new one, built on Gemini 3.6 Flash.

Nano Banana 2 had a short life as a stable product. Google's deprecation page shows its preview version arrived on February 26, 2026 and shut down on June 25. The stable version followed on May 28, 2026 and now has five months on the calendar in total.

Like a phone that gets a new chip inside but keeps the same case, the name barely changed while the engine underneath jumped from Gemini 3.1 to Gemini 3.6.

Not quite, and Google doesn't claim it is. Pro is still the heavyweight. 2.1 is the one built to be fast and cheap.

The Decoder points out that 2.1 sometimes beats Pro by a wide margin in Google's benchmarks. But Nano Banana 2 also matched Pro in those tests, and in practice Pro often still made noticeably better images.

The outlet ran its own test: a famously tricky prompt with a monkey holding a pink banana on a tiger, and a horse riding an astronaut. Nano Banana 2.1 followed it fairly closely. Side by side, Pro's image still looked more realistic, especially in colors and proportions. 2.1's horse came out looking more like a pony.

For most people, "almost as good as Pro, at a quarter of the price" is the version that ends up in everyday apps.

Google's model card reports human preference scores, called Elo ratings, from side-by-side votes. In thinking mode, 2.1 scored 1050 for overall text-to-image preference, 1048 for infographic design, 1026 for general editing and 1106 for keeping multiple characters consistent. The model's knowledge cutoff is March 2026, the card says.

Before it draws, the model can reason about the request: what goes where, what the text should say, how the pieces fit. Developers pick one of three levels: minimal, medium or high. Medium is the default.

Think of it like a sketch artist who can either start drawing right away or spend a minute planning the layout first. The Decoder notes the level affects image quality. More planning usually means a better picture and a longer wait.

It supports 1K, 2K and 4K output at what Google calls Flash-level speed.

Google says Nano Banana 2.1 is already rolling out across its products:

• Gemini app: where most people make images with Google today.
• AI Mode in Google Search: image creation right inside Search.
• Flow and Stitch: Google's tools for video and for designing app screens.
• Google Ads: for advertisers making campaign images.
• Google AI Studio and the Gemini API: for developers building their own apps.
• Gemini Enterprise: for companies.

What's next: developers on the old model have a deadline. Google's deprecation page says gemini-3.1-flash-image shuts down on October 29, 2026, and names 2.1 as the replacement. That's a little over three weeks.

The Decoder also expects the next big jump in image quality to arrive with a Pro model built on Gemini 4 Argon, Google's newest frontier model.

Google's model card is open about the limits:

• Small or long text: tiny print and long passages can still come out wrong.
• Consistency: characters don't always match the input perfectly.
• Masked edits: instruction following there is only partial.
• 3D and facts: advanced spatial reasoning and factual accuracy are limited.

On safety, Google says red teaming found similar or improved safety performance compared with Gemini 3 Flash. The model card adds that it met Google's required launch thresholds for child safety. Every image also carries a SynthID watermark, an invisible marker that tools can detect to show a picture was made by Google's AI.

• If you use the Gemini app: your images should get cleaner text and better edits without you changing anything.
• If you build apps: your per-image bill can roughly halve, but switch before October 29 and check the higher input price against your usage.
• If you need the best realism: Nano Banana Pro is still the safer pick for photo-like images.
• If you run a small business: menus, flyers and product shots just got cheaper to draft with AI. Check spelling before you print.