Nano Banana Pro

Nano Banana Pro Text-to-Image API on E2X

Nano Banana Pro is the product name Google gives gemini-3-pro-image, the top rung of its image family. We run its generation endpoint as google/nano-banana-pro/text-to-image — nothing but a prompt goes in — and we charge $0.075 per image at 1K and 2K, with 4K billed at $0.15. fal.ai wants twice that for the identical request.

Already have a picture you'd rather alter than replace? Nano Banana Pro edit-image is the same model at the same price, with an image_urls field for your references.

Where our number sat at the August 26, 2026 check:

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.

resolution is our one billing factor here. 1K and 2K cost the same $0.075 — there is no reason to render at 1K unless bandwidth matters to you — and 4K doubles it.

What the Pro tier does from a cold start

Generation from nothing is a harder problem than editing. There's no reference to anchor the composition, so everything the model gets right has to come out of its own reasoning about your sentence. Three capabilities carry most of the weight.

Typography that survives. This is the reason to pay Pro prices. The model renders styled, readable words inside the frame — headlines, price lists, menu boards, chart labels, packaging copy — across multiple languages, at sizes where cheaper models produce decorative squiggles. A poster with a real slogan on it is a single request, not a request plus an hour in a design tool.

Google Search grounding. The model can consult live search while it composes. Ask for a specific building, a specific uniform, a specific dish plated the way a specific cuisine plates it, and it has somewhere to look besides its own priors. That matters far more in text-to-image than in editing, where the source photograph already settles the question of what a thing looks like.

Real 4K, up to 4096×4096. Not an upscale bolted on afterwards. Our resolution enum is lowercase — 1k, 2k, 4k — and only 4k changes the price.

There's a fourth thing worth knowing because it shows up in your latency graphs rather than your invoices. The model reasons visually before it commits, producing interim "thought images" along the way. Google doesn't return them, and we don't charge for them. Budget around 30 seconds per request and the number stops being surprising.

API request: prompt to finished frame

prompt is the only required field. Mint a key in the dashboard, keep it behind your own backend, and post to our submit endpoint.

curl -X POST https://api.e2x.ai/v1/jobs/submit \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/nano-banana-pro/text-to-image",
    "input": {
      "prompt": "A vintage travel poster for the Faroe Islands. Screen-printed look, four flat colours, thick fog over basalt cliffs. The words VISIT FAROE set in a condensed sans serif across the bottom third.",
      "aspect_ratio": "2:3",
      "resolution": "2k"
    }
  }'

You get a job ID back. Poll it, or hand us a webhookUrl in the submit body and we'll notify you instead:

const headers = {
  Authorization: `Bearer ${process.env.E2X_API_KEY}`,
  "Content-Type": "application/json",
};

const submit = (input) =>
  fetch("https://api.e2x.ai/v1/jobs/submit", {
    method: "POST",
    headers,
    body: JSON.stringify({ model: "google/nano-banana-pro/text-to-image", input }),
  }).then((r) => r.json());

const { data } = await submit({
  prompt:
    "Isometric infographic of a coffee supply chain, six labelled stages, muted earth palette, all labels in Portuguese and legible at small size.",
  aspect_ratio: "16:9",
  resolution: "4k",
});

let job;
do {
  await new Promise((r) => setTimeout(r, 2500));
  job = await fetch(`https://api.e2x.ai/v1/jobs/${data.jobId}`, { headers })
    .then((r) => r.json());
} while (job.data.status === "pending" || job.data.status === "processing");

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

Two schema details that catch people out. Our aspect_ratio default is 9:16, so a request that omits the field comes back portrait — set it every time. And the enum is wide: ten ratios from 1:1 and 4:5 through 21:9, which covers most ad and social formats without cropping afterwards. The machine-readable spec carries the current list and the current price.

Choosing between our text-to-image models

Pro is the priciest generator in this family, and plenty of work doesn't need it:

Start at Nano Banana 2. It costs a little over half of Pro, it renders text well, and for product shots, thumbnails and editorial imagery you will rarely be able to tell which tier produced the file. Move up to Pro when the words in the image are the deliverable — a menu, an infographic, a bilingual campaign asset — or when a client needs 4096 pixels on the long edge.

One row deserves a second look: Nano Banana 2 Lite is cheaper than the legacy Nano Banana and newer than it. Google itself calls the original its legacy model and recommends Lite as the replacement. If you're still calling the old slug, that's free money on the table.

GPT Image 2 is the one genuine rival for typography, and it arrives at a different look — cooler, more literal, and able to return a transparent background, which no Gemini image model does. The remaining options live in the text-to-image category, and everything we run is on the model catalog page.

Getting the words right in the frame

Quote the string you want. Put it in quotation marks, spell it exactly, and say where it goes — "the words 'HALF PRICE TUESDAY' across the top, condensed sans, white on red." Describe the text and the model interprets; quote it and the model transcribes.

Set the type before the scene. Prompts that open with the layout and typography, then describe the imagery around them, hold their copy better than prompts that bury the text in the last clause.

Then let it work. Because this model reasons before it paints, a prompt with real structural constraints — count the elements, name the palette, fix the reading order — gets used. Vague adjectives get averaged.

Before you ship

Every image carries a SynthID watermark, Google's imperceptible provenance signal. There is no parameter to switch it off anywhere, including here. Detection tooling downstream will see it, so if a contract prohibits watermarked deliverables, sort that out first.

Check your aspect ratio assumptions if you're migrating a config from another provider. Portrait is our default, not square, and a pipeline that assumed 1:1 will silently produce 9:16 files until someone notices.

Frequently asked questions

Which Google model is Nano Banana Pro?

It's gemini-3-pro-image, the premium tier of Google's Nano Banana family. The tier is built for demanding visual work — legible in-image typography, search-grounded detail, output up to 4096×4096 — rather than for the fastest possible turnaround.

How much does Nano Banana Pro text-to-image cost on E2X?

$0.075 per image at 1K and 2K, and $0.15 at 4K. That's half of fal.ai's rate and 44% below Google's own API, as of our August 26, 2026 check. Prices move, so confirm on the live model page before you build a budget on it.

Can it generate images with readable text?

Yes, and it's the main reason to choose this tier. It renders styled headlines, labels and body copy inside the image, in multiple languages, at sizes where most image models fail. Put the wording in quotation marks in your prompt instead of paraphrasing it.

What is the difference between the text-to-image and edit-image endpoints?

Same model, same $0.075, different input. Text-to-image takes a prompt only. Edit-image additionally takes an image_urls array of references, which is what you want when an existing picture has to be changed rather than recreated.

How long does one generation take?

Around 30 seconds. The model produces interim "thought images" while it reasons about the composition — Google doesn't return them and we don't bill them, but they're where the time goes. Flash-tier models are several times faster if latency outranks quality for you.

Are Nano Banana Pro images watermarked?

Yes. Every output includes a SynthID watermark, Google's invisible marker for AI-generated content. No provider exposes a way to disable it, us included.

Is Nano Banana Pro worth it over Nano Banana 2?

Only sometimes. Nano Banana 2 costs $0.04 against $0.075 and holds up well on most photographic and illustrative work. Pay for Pro when the image contains copy someone will actually read, when you need search-grounded accuracy, or when the output has to be 4K.

Which aspect ratios are supported?

Ten of them: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9 and 21:9. Our default is 9:16, so set the field explicitly unless you want portrait output.