Nano Banana vs Nano Banana Pro, 2 and Lite: Which to Use

Our Nano Banana catalog now includes Nano Banana 2.1, and the question we get most often is not "how good is it" — it is "which one am I supposed to call". Fair question. The names do not help: a model called Lite is newer than the model with no suffix, and the one called 2 is not the one called Pro.

The short version, if you only read one paragraph: Nano Banana 2 Lite is both cheaper than the original Nano Banana and a full generation newer. That single fact kills the most common reason people stay on the legacy model. There is no budget argument for the old one anymore, and Google shuts it down on 2 October 2026 anyway.

Where the money actually goes

The existing model prices below were checked on August 26, 2026; we added Nano Banana 2.1 and its current discounted prices on October 7, 2026. All prices are per image. The legacy row is retained for migration context.

Prices and product terms can change. Check each provider's current pricing before making a purchasing decision. Google Batch or Flex pricing is not directly equivalent to a standard on-demand API request because scheduling, availability, and processing conditions differ; it is therefore excluded from this comparison. This is a scoped comparison, not a claim that E2X is the world's cheapest option in every configuration.

Read the first two rows again. The legacy model costs 31% more per image than the model Google itself recommends as its replacement. That is the whole article in two lines.

The argument that no longer works

For about a year, the honest case for staying on gemini-2.5-flash-image was price. It was the cheap one. You batched it, you ate the rough edges, you moved on.

That case is gone. Nano Banana 2 Lite is gemini-3.1-flash-lite-image, it sits at $0.0238 against the legacy model's $0.0312 — 24% less — and it is a Gemini 3 model rather than a Gemini 2.5 one. Google's own documentation calls the old one legacy and points developers at Lite for "enhanced quality, faster generation speeds, and lower API pricing". We have no reason to argue with that framing, because our price ladder says the same thing.

Then there is the calendar. Google has published a hard end date for gemini-2.5-flash-image: 2 October 2026. Not a soft deprecation. The endpoint goes away.

Pick by the job, not by the tier

Product shots for a catalog

Start with Nano Banana 2 Lite edit-image. Background swaps, surface changes, relighting a packshot against a plain grey sweep — Lite handles that class of work at 1K, which is more than a product grid needs, and it takes up to 14 reference images in a single call.

Move up to Nano Banana 2 when one prompt has to juggle several objects at once. That is the clearest quality gap between the two: multi-object instruction-following. If your prompt reads "put the bottle left of the box, tilt the label toward camera, add a second unit behind", Lite starts dropping clauses and NB2 does not. Also move up if a zoom view needs 2K or 4K, since Lite has no resolution parameter at all.

Social variants of one asset

Lite, and it is not close. Its aspect ratio list goes well past the usual set — 1:4, 4:1, 1:8, 8:1 and an auto mode are all there, so banner rails and skyscrapers come out of the same endpoint as your square. Google says the model can produce an image in as little as four seconds.

One operational note: our published ETA on the Lite endpoints is 60 seconds, not four. That figure covers queueing and delivery on our side, not the model's raw inference time. Budget your timeouts against ours, not Google's.

Anything with words baked into the image

This is where the generations actually separate. The legacy Nano Banana is weak at in-image text, and we are not going to pretend otherwise. Google only claims advanced text rendering starting with the Gemini 3 models, which the old one is not. Infographics, packaging mockups, menu boards, anything with a headline in it: the old model turns body copy into decorative marks.

Both Nano Banana 2 and Lite render legible in-image text. Published comparisons put the two level on that. It is not a test we ran, so if the copy is the point of the asset, sample both before you commit. For a caption or a product label, the cheap one is fine. Go to Nano Banana Pro when the layout is dense and multilingual: it will re-typeset an English infographic into Spanish without redrawing the chart underneath it.

If your job is mask-driven instead — you want a specific region replaced and everything outside it untouched — none of these do that. GPT Image 2 edit-image does, and it does transparent backgrounds too. We bill it at $0.0525 per image at 1K. One thing to know before you plan cheap draft passes there: its quality field saves nothing, since low, medium and high all land on that same figure. Resolution is the only lever on spend. Different vendor, different strengths.

High-volume batch runs

Lite. Run the arithmetic on 10,000 images a month:

• Nano Banana 2 Lite: $238
• Nano Banana (legacy): $312
• Nano Banana 2: $400
• Nano Banana Pro: $750

The Pro row is the one worth staring at. Three times Lite, for a model most batch pipelines are not using the strengths of. We would rather point you down a tier than bill you for one your pipeline never touches.

Character or product consistency across a set

Both Nano Banana 2 and Nano Banana 2 Lite accept up to 14 reference images, which is the number that matters here, and they tie on reference consistency. Start with Lite.

Pro earns its price when those references carry different jobs. Its fourteen slots are not interchangeable — Google caps style references at three, object references at six, and lets five carry a face that has to survive intact. Push a person, a product and a brand palette through together and it is the role assignment that stops the model from reading your palette as a subject. The flash tiers take the same fourteen images in a single call — just as a flat list, with nothing marking which attachment was a colour reference and which was a face.

4K for print

Lite is out, since 1K is all it makes. That leaves three:

• Nano Banana 2: Most affordable 4K at $0.08 per generation.
• Nano Banana 2.1: Our lowest global Nano Banana 2.1 prices are $0.03 (1K), $0.04 (2K), and $0.09 (4K), with 15 aspect ratios including 1:8 and 8:1. These are 50% below our $0.06 / $0.08 / $0.18 list prices. The discount stays in place unless Google changes its pricing; it is not a temporary launch offer. Read the announcement.
• Nano Banana Pro: Premium tier at $0.15 for search-grounded detail, role-separated references, and intricate typography.

Moving off the legacy model

Treat 2 October 2026 like any other dated dependency going away: a scheduled incident with a known fix. If the switch is still on your backlog the week Google pulls it, every image call in your product fails at once.

The fix is one string. For text-to-image it is genuinely the whole diff:

 {
-  "model": "google/nano-banana/text-to-image",
+  "model": "google/nano-banana-2-lite/text-to-image",
   "input": {
-    "prompt": "studio shot of a walnut desk lamp on a plain grey sweep"
+    "prompt": "studio shot of a walnut desk lamp on a plain grey sweep",
+    "aspect_ratio": "9:16"
   }
 }

That extra line is the one thing that will bite you. Our legacy endpoints default to 9:16; the Lite endpoints default to 1:1. If your code never sent aspect_ratio because the old default happened to match your layout, pin it explicitly before you cut over or every asset silently becomes a square.

Two other differences to check while you are in there. Lite has no resolution parameter — 1K is the only output it produces, so drop the field if you were sending it to the legacy edit endpoint. And Lite has no search grounding, where Nano Banana 2 does — if you were weighing a two-tier jump anyway, that is the argument for NB2.

Here is the whole call, end to end, against the Lite edit endpoint:

curl -X POST https://api.e2x.ai/v1/jobs/submit \
  -H "Authorization: Bearer $E2X_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/nano-banana-2-lite/edit-image",
    "input": {
      "prompt": "Drop the shirt onto a plain bone-white background. Keep the fabric texture and the embroidered logo exactly as they are.",
      "image_urls": ["https://example.com/shirt-on-hanger.jpg"],
      "aspect_ratio": "4:5"
    },
    "webhookUrl": "https://your-app.example.com/hooks/e2x"
  }'

Send webhookUrl and we call you when the job lands. Skip it and you poll instead — submit returns a job ID, and you read the status until it settles:

const auth = { Authorization: `Bearer ${process.env.E2X_API_KEY}` };

async function waitForJob(jobId, deadlineMs = 120_000) {
  const stopAt = Date.now() + deadlineMs;

  while (Date.now() < stopAt) {
    const res = await fetch(`https://api.e2x.ai/v1/jobs/${jobId}`, { headers: auth });
    const { data } = await res.json();

    if (data.status === "completed") return data.outputs[0].url;
    if (data.status === "failed") throw new Error(data.error?.message ?? "job failed");

    await new Promise((r) => setTimeout(r, 2_000));
  }

  throw new Error(`job ${jobId} did not settle in time`);
}

Status moves pending to processing to completed, or stops at failed or cancelled. Same contract on all four models, which is the point of putting one API in front of them.

Before you ship

Every output from all four carries a SynthID watermark. Google embeds it invisibly, on every image, with no opt-out exposed to any provider — ourselves included. Settle that question with your legal or brand team before the integration goes in, not after a client asks why an asset trips their detection tooling.

Check your aspect ratio defaults against ours whenever you move between families, not just on the legacy migration. And if you are still deciding, the two category pages are the fastest way to see everything side by side: text-to-image and image-to-image, or the full model catalog.

Frequently asked questions

What is the difference between Nano Banana and Nano Banana Pro?

Nano Banana is Google's original gemini-2.5-flash-image, a 1K-only flash-tier model that Google now labels legacy and retires on 2 October 2026. Nano Banana Pro is gemini-3-pro-image, a premium model built for multi-reference composition, legible in-image text and output up to 4K. They are two generations apart, and we charge $0.0312 and $0.075 per image respectively.

Which Nano Banana model should I use?

For most work, Nano Banana 2 Lite at $0.0238 per image. Step up to Nano Banana 2 when one prompt has to control several objects at once, or when 1K is not enough — it bills $0.04 at 1K, and the ladder runs $0.06 for 2K, $0.08 for 4K. Pro at $0.075 is worth it for dense multilingual text layouts and role-separated reference composites, and it is overkill for everything else.

Are Nano Banana 2 and Nano Banana Pro the same model?

They are not, and you will find catalogs that label them as if they were. One is built on gemini-3.1-flash-image, the other on gemini-3-pro-image, and we bill them at $0.04 and $0.075 per 1K image respectively. Separate slugs on our side, deliberately.

Why is Nano Banana 2 Lite cheaper than the older Nano Banana?

Because the newer architecture is cheaper for Google to serve, and that flows through to our pricing. Lite sits at $0.0238 against the legacy model's $0.0312 as of our August 26, 2026 check. It is the unusual case where the upgrade path also lowers your bill.

How hard is it to migrate off the legacy Nano Banana?

For text-to-image it is one string: swap google/nano-banana/text-to-image for google/nano-banana-2-lite/text-to-image. Pin aspect_ratio explicitly at the same time, because the legacy default is 9:16 and Lite's is 1:1. On the edit endpoint, also drop any resolution field, since Lite produces 1K only.

Can I remove the SynthID watermark from the output?

No. Google embeds SynthID in every image these models produce, at every tier, and there is no API parameter on our side or anyone else's to disable it. Plan for it in any workflow where provenance metadata or AI-detection labelling would be a problem.

Do all four models use the same API request format?

Yes. Every model in our catalog is submitted to POST https://api.e2x.ai/v1/jobs/submit with a bearer token, a model slug and an input object, and you either poll the job or receive a webhook. Swapping models means changing the slug and checking the parameters that model actually accepts.