Nano Banana Pro vs GPT Image 2.5: a test on identical prompts

01How did we run the test?

• One prompt for both models. A drink poster, a three-step infographic, and a headphones product card, each with lettering. Both models got the same prompt, word for word.

• Same settings. Both ran at 1K through the Rendergrid API, at 2:3, 3:2, and 1:1.

• One attempt per model. We show that one result, with no picking of the best of several, and we score nothing.

02Do both models set the poster’s lettering correctly?

Yes. The prompt asks for “Summer afternoon” above the bottle, “Lemonade” beneath it, and nothing else in writing. Both models wrote both phrases exactly.

One prompt for both models

Design a portrait drink poster around a clear lemonade bottle on a pale wooden table. Include two lemon slices and a mint sprig. Warm light enters from the side; keep the cream background quiet and leave generous margins. Set “Summer afternoon” above the bottle and “Lemonade” beneath it in plain, readable type. Show the entire bottle. Include only the specified lettering, with no logos, people, or watermarks.

1K · 2:3 · one attempt per modelCopy prompt

The scenes differ. Nano Banana Pro drew an open bottle of amber-colored drink, hard-edged shadows, and a dark band down the left of the wall, and some lemon and mint are cut off by the frame’s edges. GPT Image 2.5 drew a capped bottle of pale yellow lemonade and soft, leaf-shaped shadows on the wall, with every prop inside the frame. Both put more lemon and mint on the table than the prompt names.

How to write lettering in a prompt

In this test, both models followed the three habits of the prompt:

1. Put the exact words in quotation marks: “Summer afternoon” and “Lemonade”.
2. Give each phrase a place: “above the bottle” and “beneath it”.
3. Bar everything else: “Include only the specified lettering”. None of the six images has stray text.

More lettering prompts are in Text in AI images: how to get readable lettering.

03How do the models draw a three-step infographic?

Both labeled the pots “Pot”, “Soil”, and “Seedling” exactly and set three evenly spaced pots with arrows from left to right. Nano Banana Pro drew a flat, outlined illustration with thin arrows; the base of its first pot is ragged, a flaw the other two pots don’t share. GPT Image 2.5 drew softly shaded pots with smooth bases, and its arrows are short and heavy, where the prompt asked for thin ones. In both, the side margins are narrower than the top and bottom ones, where the prompt asked for wide outer margins.

One prompt for both models

Make a landscape infographic showing a seedling being potted. Three evenly spaced drawings on cream show an empty clay pot, a pot filled with soil, then the pot with a small green seedling. Thin arrows connect them from left to right. Label the drawings “Pot”, “Soil”, and “Seedling” in large, simple lettering. Keep wide outer margins. Include only the specified lettering, with no logos, people, or watermarks.

1K · 3:2 · one attempt per modelCopy prompt

The prompt asks for “large, simple lettering”, and both models set the labels large. Name the size of the lettering in the prompt.

04How do the models draw a product card?

Both labeled “Headband”, “Ear cushions”, and “Ear cups”, kept every label off the headphones, and showed the whole product at a three-quarter angle on light gray. Nano Banana Pro’s headphones hover above a separate oval shadow, where the prompt asked for a soft contact shadow, and its leader lines are thin horizontal rules. GPT Image 2.5’s leader lines are angled and end in small dots on their parts, and its headphones fill more of the frame, so the margins are tighter: the “Ear cushions” label starts about 5% from the left edge, where the prompt asked for generous clear margins.

One prompt for both models

Create a square product card for ivory over-ear headphones, fully visible at a three-quarter angle on light gray. Use three fine leader lines to identify visible parts: “Headband”, “Ear cushions”, “Ear cups”. Place the labels around the product without overlapping it. Add a soft contact shadow and generous clear margins; keep the typography understated. Include only the specified lettering, with no logos, people, or watermarks.

1K · 1:1 · one attempt per modelCopy prompt

More prompts for product cards are in product card prompts for Nano Banana Pro.

05How much does an image cost on each model?

A Nano Banana Pro image costs $0.04 at 1K, $0.04 at 2K, and $0.07 at 4K. A GPT Image 2.5 image costs $0.025 at 1K, its only resolution. The amount is charged when a request is accepted, and a failed request is refunded automatically.

If resolution is left out, an image is rendered and billed at the model’s default: 2K for Nano Banana Pro, 1K for GPT Image 2.5. Send "resolution": "1K" to repeat this test.

Official prices from the makers

Google publishes a price per image for Nano Banana Pro, so it can sit next to Rendergrid’s. OpenAI’s figure for GPT Image 2.5 comes from its image-generation cost calculator, at high quality and a square size. Rendergrid has no quality field, and “high” only describes that line.

All models and resolutions are on the pricing page.

06How long does an image take?

There is no promised time: it depends on the model and the load. Poll for the status no more often than once every 5 seconds, and limit how long you wait.

An average describes past requests across a model’s traffic, not these six images and not your next request.

07How do the two models compare in numbers?

08How do you repeat the test through the API?

Both models take the same request; only the model field changes. The sample sends this test’s poster prompt to each model and then reads the status. The models available to your account are listed by GET /models.

# Requires bash, curl, and uuidgen. One idempotency key per request.
IDEMPOTENCY_KEY="$(uuidgen)" || exit 1
curl --fail-with-body --max-time 30 \
  -X POST https://api.rendergrid.io/api/public/v1/images/generate \
  -H "Authorization: Bearer $RENDERGRID_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Idempotency-Key: $IDEMPOTENCY_KEY" \
  -d '{
    "model": "nano-banana-pro",
    "prompt": "Design a portrait drink poster around a clear lemonade bottle on a pale wooden table. Include two lemon slices and a mint sprig. Warm light enters from the side; keep the cream background quiet and leave generous margins. Set “Summer afternoon” above the bottle and “Lemonade” beneath it in plain, readable type. Show the entire bottle. Include only the specified lettering, with no logos, people, or watermarks.",
    "resolution": "1K",
    "aspect_ratio": "2:3"
  }'

# For GPT Image 2.5, send the same request with "model": "gpt-image-2.5"
# and a new idempotency key. Then read each status with the id from its response:
CREATION_ID="5b0214ae-9e39-460a-820a-f6ae3e24400b"  # the ID from the response
curl --fail-with-body --max-time 20 \
  "https://api.rendergrid.io/api/public/v1/creations/$CREATION_ID" \
  -H "Authorization: Bearer $RENDERGRID_API_KEY"

import os
import time
import uuid

import requests

API = "https://api.rendergrid.io/api/public/v1"
KEY = os.environ["RENDERGRID_API_KEY"]
PROMPT = "Design a portrait drink poster around a clear lemonade bottle on a pale wooden table. Include two lemon slices and a mint sprig. Warm light enters from the side; keep the cream background quiet and leave generous margins. Set “Summer afternoon” above the bottle and “Lemonade” beneath it in plain, readable type. Show the entire bottle. Include only the specified lettering, with no logos, people, or watermarks."

for model in ("nano-banana-pro", "gpt-image-2.5"):
    response = requests.post(
        f"{API}/images/generate",
        headers={
            "Authorization": f"Bearer {KEY}",
            # a new key for every new request
            "X-Idempotency-Key": str(uuid.uuid4()),
        },
        json={
            "model": model,
            "prompt": PROMPT,
            "resolution": "1K",
            "aspect_ratio": "2:3",
        },
        timeout=(10, 120),
    )
    response.raise_for_status()
    creation = response.json()

    while True:
        poll = requests.get(
            f"{API}/creations/{creation['id']}",
            headers={"Authorization": f"Bearer {KEY}"},
            timeout=(5, 20),
        )
        poll.raise_for_status()
        status = poll.json()
        if status["status"] in ("completed", "failed"):
            break
        time.sleep(5)  # do not ask more often than once every 5 seconds

    if status["status"] == "completed":
        print(model, status["result_urls"])
    else:
        print(model, status["error"], status["failure"])

// Node.js 22+, .mjs file
import { randomUUID } from "node:crypto";

const API = "https://api.rendergrid.io/api/public/v1";
const key = process.env.RENDERGRID_API_KEY;
const prompt =
  "Design a portrait drink poster around a clear lemonade bottle on a pale wooden table. Include two lemon slices and a mint sprig. Warm light enters from the side; keep the cream background quiet and leave generous margins. Set “Summer afternoon” above the bottle and “Lemonade” beneath it in plain, readable type. Show the entire bottle. Include only the specified lettering, with no logos, people, or watermarks.";
const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));

for (const model of ["nano-banana-pro", "gpt-image-2.5"]) {
  const response = await fetch(`${API}/images/generate`, {
    method: "POST",
    headers: {
      Authorization: `Bearer ${key}`,
      "Content-Type": "application/json",
      // a new key for every new request
      "X-Idempotency-Key": randomUUID(),
    },
    body: JSON.stringify({ model, prompt, resolution: "1K", aspect_ratio: "2:3" }),
    signal: AbortSignal.timeout(120_000),
  });
  if (response.status !== 202) {
    throw new Error(`HTTP ${response.status}: ${await response.text()}`);
  }
  const creation = await response.json();

  let status;
  do {
    // do not ask more often than once every 5 seconds
    await sleep(5000);
    const poll = await fetch(`${API}/creations/${creation.id}`, {
      headers: { Authorization: `Bearer ${key}` },
      signal: AbortSignal.timeout(20_000),
    });
    if (!poll.ok) throw new Error(`HTTP ${poll.status}: ${await poll.text()}`);
    status = await poll.json();
  } while (["queued", "processing"].includes(status.status));

  console.log(model, status.status, status.result_urls ?? status.error);
}

{
  "id": "5b0214ae-9e39-460a-820a-f6ae3e24400b",
  "status": "queued",
  "task_id": "5b0214ae-9e39-460a-820a-f6ae3e24400b",
  "cost": 0.04
}

{
  "id": "5b0214ae-9e39-460a-820a-f6ae3e24400b",
  "status": "completed",
  "result_urls": ["https://cdn.rendergrid.io/images/example/result.jpg"]
}

Pass webhook_url and Rendergrid tells you when the task becomes completed or failed; keep polling as a fallback. A failed request is free: the amount goes straight back to your balance.

Start with the quickstart or the Nano Banana Pro quickstart. Each model’s parameters are on its page: Nano Banana Pro and GPT Image 2.5.

09Which model should you pick for which task?

• Images at 2K or 4KGNano Banana ProOnly it has 2K and 4K.
• A poster with every prop inside the frameOGPT Image 2.5In this test its poster did; some lemon and mint in Nano Banana Pro’s were cut off.
• A diagram with thin arrowsGNano Banana ProIn this test its arrows were thin; GPT Image 2.5’s were heavier.
• A product shot with a shadow under the productOGPT Image 2.5In this test a soft shadow sat beneath its headphones, while Nano Banana Pro’s hovered above a separate shadow.

10Where does this comparison not fit?

• Three prompts are not a benchmark. On your own prompts the images can look different, so check both models on them.
• One attempt per model. We did not measure how two attempts of one prompt differ. Another attempt is a new request with a new idempotency key, charged on its own.
• Only 1K. GPT Image 2.5 has no 2K or 4K, so the larger tiers of Nano Banana Pro are outside this test.
• Not everything was tested. Edits from reference images and long texts were not part of it.
• Prices and waits are not from the test. Prices come from Rendergrid’s price list, and an average wait, where measured, covers a model’s requests over the last 7 days.

11Questions and answers

12Sources

1. Google AI for Developers: API pricingGoogle’s official price for Nano Banana Pro: standard rate, without batch mode. Checked on October 26, 2026.
2. OpenAI: image generation guideOpenAI’s cost calculator for GPT Image 2.5: high quality, a square size. Checked on October 26, 2026.
3. Rendergrid pricingThe price per image for each model and resolution.
4. Rendergrid modelsThe maker, resolutions, and aspect ratios of each model.
5. Rendergrid documentation: tasks and statusesPolling, the statuses, and the waiting time.
6. Rendergrid test, six imagesThe prompt and settings sit beside each pair of images.
7. Rendergrid statisticsWhere this article shows an average wait, it covers the model’s requests over the last 7 days.