If you have opened the Gemini app recently or read a headline about Google Earth, you have probably met the phrase “Nano Banana” and assumed it was a meme, a third-party app, or a typo. It is none of those — it is Google’s official brand name for the image models inside Gemini, a codename that leaked, went viral, and got promoted to a product name.

Quick answer

Where the silly name came from

In late July 2025, Google DeepMind needed a codename to test an unreleased image model anonymously on LMArena, where users compare two unidentified models and vote. Google’s own account, published in January 2026, is that product manager Naina Raisinghani suggested “Nano Banana” at 2:30am, mashing together two of her personal nicknames — “Naina Banana” from friends, and “Nano” because she is short and likes computers. Another PM’s reaction, per Google’s write-up: “That’s completely nonsensical.”

Users quickly noticed it was far better at editing than anything else in the rotation, particularly at keeping a person’s face recognisably the same. By the public launch on 26 August 2025 the nickname had more brand equity than the model number, so Google leaned in — and named the successor Nano Banana Pro. The codename is now the brand; the Gemini model IDs are the technical names underneath.

The four models called Nano Banana in 2026

Google’s developer documentation lists four distinct models under the name:

• Nano Banana 2 (gemini-3.1-flash-image) — 26 February 2026. Google’s generalist workhorse: speed plus 4K output, world knowledge, text rendering and good multi-reference handling. The default in most consumer surfaces.
• Nano Banana Pro (gemini-3-pro-image) — 20 November 2025, built on Gemini 3 Pro. The premium option for complex compositions, brand consistency, localisation and demanding text rendering.
• Nano Banana 2 Lite (gemini-3.1-flash-lite-image) — 30 June 2026. Fastest and cheapest, for high-volume drafting; Google says around four seconds per image. Limited to 1K, and not built for multi-reference or multi-turn editing.
• Nano Banana (gemini-2.5-flash-image) — the original, now labelled legacy. Google recommends developers move to Nano Banana 2 Lite.

If you build on Google’s image stack: the older Imagen line is being retired in favour of these models, with official migration guidance pointing developers at Nano Banana. [Insert live content] confirm whether the Imagen shutdown (listed as “as early as 17 August 2026”) has taken effect

How Nano Banana differs from other image models

Most image generators are dedicated systems: you prompt, you get a picture, and editing means running it back through with a mask. Nano Banana is native to Gemini — image generation is a capability of the language model, not a service it calls. Three practical differences follow:

• You edit conversationally. “Now make it night.” “Move the logo left.” “Same character, different angle.” No mask tool, no inpainting brush — describe the change and the model applies it, leaving the rest alone. This is what made the original go viral.
• It knows things. Sitting on Gemini’s reasoning and world knowledge, it produces content-bearing images: a labelled diagram, a recipe card, an infographic built from a document you supplied. Pro and Nano Banana 2 can also ground against Google Search, so an image can reflect real-time information.
• It renders text properly. Illegible text was the classic giveaway in AI images. Google’s pitch for the Gemini 3 generation is accurate, multilingual type — posters, mockups, packaging, and translating text inside an uploaded image.

The trade-off: it is a generalist. For a distinctive house style, dedicated art models still have a personality advantage. Nano Banana’s strengths are accuracy, consistency and editability.

How to use Nano Banana

There is no Nano Banana app — anything calling itself one is a third-party wrapper. The genuine routes:

• Gemini app (web, Android, iOS) — the main consumer route. Ask for an image, or use the create-image control. Free accounts get a limited allowance; AI Plus, Pro and Ultra subscribers get more. Pro access runs off your Gemini 3 Pro quota; exhaust it and the app falls back to the faster model rather than stopping.
• Google Search (AI Mode and Lens) — image generation inside search, now well beyond the original English/US rollout.
• Google AI Studio and the Gemini API — the developer route: pick a model ID, set resolution and aspect ratio, pass up to 14 reference images.
• Slides, Vids, Flow, Google Ads, NotebookLM, Vertex AI — specialist surfaces for decks, filmmaking frames, ad creative, notebook visuals and enterprise deployment.

API pricing is per generated image, by model and output resolution, with a batch discount. [Insert live content] per-image API prices for all four models, plus current daily app quotas by tier Google does not publish fixed app quotas, so treat figures quoted on third-party sites with suspicion.

Prompt patterns that actually work

Nano Banana rewards specificity far more than keyword-stuffing. These patterns come from Google’s own prompting guidance.

1. Describe a scene, not a subject. Weak: “a coral reef.” Strong: shot type, subject, setting, lighting, camera angle, lens — and the aspect ratio, which is a real parameter rather than a suggestion. Google’s template is roughly A photorealistic [shot type] of a [subject] in a [setting]. [Description of the light]. Shot from a [angle] with a [lens].

2. Put text you want rendered in quotes. Write the words exactly, then say where they go and how they look: a poster with the headline “Late Shift” in a heavy condensed sans-serif, centred, cream on deep green. Vague instructions produce plausible-looking gibberish; explicit strings produce the string.

3. Edit by describing the delta. Say only what changes — “remove the bin from the left of the frame and match the pavement texture” beats re-describing the scene. The model preserves what you did not mention.

4. Use reference images for consistency. Up to 14 references per request, with separate budgets for objects, characters and style — the intended way to keep a character or brand look stable across a set. Stay in one thread as you iterate so the model keeps what you have already fixed.

What Nano Banana Pro adds

Pro is what most people mean when they search for a “nano banana pro ai image generator”. The differences are concrete:

• Resolution. 1K, 2K and 4K output, versus the Lite model’s 1K ceiling.
• Composition budget. Up to 14 reference images, with consistency held across up to five people in one frame — the feature behind group shots and storyboards that do not drift.
• Studio controls. Localised editing of a selected region, camera angle, focus and depth of field, colour grading, and relighting (day to night, chiaroscuro, bokeh).
• Text and localisation. The strongest text rendering in the family, including translating text inside an uploaded image while leaving the design intact.
• Watermark handling. Free and AI Pro tier images carry a visible Gemini sparkle mark; Google removes it for AI Ultra subscribers and in AI Studio, reasoning that professional work needs a clean canvas. The invisible SynthID watermark stays regardless.

For quick edits and drafts, Nano Banana 2 is faster with a more generous quota. For a poster with real typography, a multi-character scene, or anything going to print, Pro is the one.

Provenance: SynthID and C2PA

Every image these models generate carries SynthID, Google’s imperceptible watermark, embedded in the pixel data rather than file metadata. Since November 2025 you can upload an image into the Gemini app and ask whether Google AI made it; Google reported that feature had been used over 20 million times by February 2026. Google is also pairing SynthID with interoperable C2PA Content Credentials, which record not just whether AI was involved but how.

These are legitimate transparency signals. They are not complete — a mark tells you an image came from a generator, but its absence does not prove one is real — yet they are the most practical check available, and free to use.

Honest limits

Confident factual errors. The infographic capability is the headline feature and the biggest trap: a model that renders a clean, professional diagram with legible labels can just as cleanly render one with wrong labels. Check anything with numbers, dates, names or measurements.

Quotas are opaque and they move. Daily allowances have been adjusted repeatedly and Google publishes no firm numbers. If your workflow depends on volume, the API’s published pricing is more predictable.

Realism is the risk, not just the feature. On 30 July 2026 Google added Nano Banana editing to Google Earth, letting users alter genuine satellite and aerial imagery from a text prompt. Within roughly two days it was withdrawn, after fabricated scenes — a refugee column at a border, a damaged hospital, a nuclear facility — circulated as though real. Google said it would restore the feature only with stronger guardrails. The better a model gets at photorealism, the more consequential a fake becomes. [Insert live content] whether the Google Earth feature has since returned

It is not a video model. Nano Banana generates and edits stills; video is Veo’s job — see our Sora 2 vs Veo comparison.

So, is it worth using?

For editing existing photos and for images that carry information, Nano Banana is currently the most useful thing in the category, and being free at low volume inside an app most people already have removes nearly all the friction. For pure artistic generation it is competent rather than distinctive. It is also a capability inside a broader assistant, not a standalone tool — which matters if you are choosing between platforms: see our Gemini vs ChatGPT comparison.