Nano Banana

Overview

Nano Banana is Google's family of native image generation models inside Gemini — an internal codename that escaped and became the product name people actually search for. It now covers four models across two generations, and the naming does more to obscure the lineup than to explain it.

The family in August 2026:

The thing worth knowing before choosing between them is that price ordering and quality ordering do not match. On Artificial Analysis's blind-vote text-to-image arena, checked 2 August 2026, Nano Banana 2 Lite sits at rank 4 with 1,263 Elo and Nano Banana 2 at rank 6 with 1,262 — while Nano Banana Pro, at four times the price of Lite, sits at rank 9 with 1,225.

"Pro" here does not mean better output. It means a different capability set: advanced text rendering and localisation, and interleaved text-and-image output. If you are not using those, you are paying a premium for a lower preference score.

The family competes against Seedream 5.0, Qwen-Image 2.0, HiDream-O1-Image and OpenAI's GPT Image line, which currently leads that arena at 1,339.

Capabilities

• Text-to-image with thinking mode. All four models reason about complex prompts before generating rather than mapping the prompt straight to pixels.
• Reference images, with different budgets per model. Nano Banana 2 accepts up to 10 character references plus 4 style references; 2 Lite accepts up to 14 object references; Pro accepts up to 6 character references. These ceilings, not the price, are usually what decides which model fits a pipeline.
• Google Search grounding. Nano Banana 2 can ground generation in image search results — useful when the prompt names a real place, product or person and accuracy matters. 2 Lite does not have it.
• Multi-turn editing. Available on Nano Banana 2 and Pro; not on 2 Lite, which is single-shot.
• Video-to-image. Nano Banana 2 can generate stills from video input.
• Advanced text rendering and localisation. Pro's distinguishing feature. Legible, correctly laid-out text inside an image remains the hardest thing image models do, and this is what the premium buys.
• Interleaved text-and-image output. Pro only — the model returns prose and images in one response rather than an image alone.
• SynthID watermarking. Applied by every model, with no opt-out.

Technical Specifications

• Model IDs: gemini-3-pro-image, gemini-3.1-flash-image, gemini-3.1-flash-lite-image, gemini-2.5-flash-image
• Resolutions, Nano Banana 2: 512px, 1K, 2K, 4K
• Resolutions, Nano Banana Pro: up to 4K
• Resolutions, 2 Lite: 1K only
• Aspect ratios: 1:1, 3:2, 2:3, 16:9 and others
• Watermark: SynthID on all outputs
• Thinking mode: all models
• Grounding: Google Search with image search — Nano Banana 2 only
• Multi-turn editing: Nano Banana 2 and Pro
• Access: Gemini API, Vertex AI, the Gemini app, and Google Antigravity, where Nano Banana 2 handles inline image generation and is not user-selectable

Use Cases

• General production image work. Nano Banana 2 is the default choice: it is the only model in the family spanning 512px to 4K, it grounds in search, and it supports multi-turn editing. Most pipelines should start here and only move if they hit a specific wall.
• High-volume generation at 1K. 2 Lite at $0.0336 per image, with a top-5 arena placement. At roughly 30 images per dollar it is the cheapest credible option Google offers, and the quality trade-off against its more expensive siblings is smaller than the price gap implies.
• Images containing text. Posters, UI mockups, signage, anything with a headline in it. This is the one job where Pro's premium is defensible, and where most image models still fail visibly.
• Localised creative. Pro's text localisation matters when the same layout ships in several languages and the type has to be right in each.
• Character consistency across a set. The reference-image budgets decide feasibility: 10 character references on Nano Banana 2 is the widest in the family.
• Object-heavy composition. 2 Lite's 14 object references is the highest ceiling of the four, which is an odd but real advantage for the cheapest model.
• Illustrated documents. Pro's interleaved text-and-image output produces prose and figures in one pass, rather than requiring separate calls and manual assembly.

Performance / Benchmarks

Artificial Analysis text-to-image arena, checked 2 August 2026. Elo derived from blind user votes on pairs of images generated from the same prompt.

Two things follow from this table.

The internal ordering is inverted against price. Lite at $0.0336 outranks Pro at $0.134–$0.24. A one-point Elo gap between Lite and Nano Banana 2 is noise; the 38-point gap down to Pro is not. Choose Pro for its capability list, never for expected image quality.

Google does not lead this board. GPT Image 2 is 76 Elo ahead of Google's best entry. Nano Banana's case against it is price — $0.0336 per 1K image is an order of magnitude below what frontier image APIs typically charge — and integration with the rest of Gemini, not raw preference.

As with any arena: voters are judging aesthetics and prompt adherence on isolated images. They are not judging text rendering, reference-image ceilings, grounding, or watermark policy, all of which decide production suitability more often than a preference score does.

Limitations

• SynthID is mandatory. Every output carries the watermark and there is no opt-out. For most uses this is irrelevant; for some client work it is disqualifying, and it should be checked before a pipeline is built.
• The naming is genuinely confusing. "Nano Banana 2" and "Nano Banana Pro" are different generations — 3.1 Flash and 3 Pro respectively — so the newer model is the one without "Pro" in its name. Third-party write-ups routinely conflate them, and pricing articles quote the wrong tier.
• 2 Lite is single-shot and 1K only. No multi-turn editing, no grounding, no resolution above 1K. Its arena placement makes it tempting for everything; its feature list does not support that.
• Pro underperforms its price on preference. Rank 9 at four times the cost of rank 4 within the same family.
• The legacy model is still callable. gemini-2.5-flash-image remains available at $0.039 — more expensive than 2 Lite and older. There is no reason to start new work on it, and Google recommends migrating.
• Paid tier for API access. Consumer use runs through Google AI subscriptions with credit metering rather than per-image pricing.

Pricing & Access

Standard per-image rates from Google's pricing page:

Batch pricing halves most of it: Nano Banana 2 drops to $0.022 / $0.034 / $0.050 / $0.076 across its four resolutions, and Pro to $0.067 at 1K/2K and $0.12 at 4K. 2 Lite is $0.0336 on both standard and batch — there is no batch discount because there is little left to discount.

Pro also has a priority tier billed by tokens at $216.00 per million rather than per image.

The comparison that matters: Nano Banana 2 at 4K costs $0.151, which is more than Pro at 1K ($0.134) and less than Pro at 4K ($0.24). If you want maximum resolution at minimum cost, Nano Banana 2 is the answer, and it also scores higher on the arena.

Access is through the Gemini API and Vertex AI on the paid tier, the Gemini app for consumers, and inside Google Antigravity for inline generation.

Ecosystem & Tools

• Gemini API — primary developer route for all four models
• Vertex AI — enterprise path with batch and priority tiers
• Gemini app — consumer access, credit-metered
• Google Antigravity — uses Nano Banana 2 for UI mockups and diagrams; the model is fixed and not user-selectable
• Google Gemini — the assistant surface these models sit behind

Community & Resources

• Gemini image generation docs — model IDs, reference-image limits, capability matrix
• Gemini API pricing — the authoritative per-image rates
• Artificial Analysis Image Arena — the leaderboard cited above
• Gemini image models on DeepMind — model family page