Nano Banana Pro API: 3 Model IDs Compared, from $0.03 per Image

If you see three model ids for Nano Banana Pro (Google's Gemini 3 Pro Image) on an API gateway, the short answer is: use gemini-3-pro-image by default, keep gemini-3-pro-image-preview only if it is already in your code and you want $0.068 instead of $0.08 at 1K and 2K, and send work that can wait to gemini-3-pro-image-gemini, a flat $0.03 per image that is much slower with reference images. Plan for 30 to 90 seconds per image. The rest of this post backs each point with numbers.

Disclosure up front: the numbers below come from apimodels.app, which is the API I work on. apimodels.app is a multi-model API gateway: one API key and one OpenAI-compatible endpoint for image, video, audio and language models. Latency and failure figures are from our production traffic over the last 7 days (sample size not disclosed), rounded to whole seconds and whole percentages.

Three ids, one model: which to use

All three ids are Nano Banana Pro with the same request body and response shape. They differ in price, refusal rate and speed.

  • gemini-3-pro-image is the default. It is the id Google itself lists for the model, it had the lowest refusal share of the three (about 2%), and it has enough traffic for us to quote timings at 1K, 2K and 4K. It costs $0.08 per image at 1K or 2K and $0.13 at 4K.
  • gemini-3-pro-image-preview costs $0.068 at 1K or 2K and the same $0.13 at 4K. At 2K it ran at about the same speed as the default id, and about 4% of its prompts were refused. Google retired this id on its own API on June 25, 2026 in favor of gemini-3-pro-image; it keeps working here for existing integrations, but new code should use the default id.
  • gemini-3-pro-image-gemini is a flat $0.03 at any resolution. Text-to-image requests ran close to the default id (median 45 s), but with reference images the median was about 100 s and the p90 passed three minutes. About 4% of its prompts were refused. It suits overnight batches, drafts and anything nobody is waiting on.

Rule of thumb: if a person is watching a spinner, use gemini-3-pro-image; if a queue is waiting, gemini-3-pro-image-gemini does the job for less than half the price.

What a request looks like

Requests are asynchronous by default: you create a task, get a task id back at once, then poll it or receive a webhook at callback_url. Example below (written against our docs, not run; set APIMODELS_API_KEY first):

import os
import time
import requests

API = "https://api.apimodels.app/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['APIMODELS_API_KEY']}"}

def generate(model, prompt, image_urls=None, resolution="2k", aspect_ratio="16:9"):
    body = {"model": model, "prompt": prompt,
            "resolution": resolution, "aspect_ratio": aspect_ratio}
    if image_urls:
        body["image_urls"] = image_urls  # public URLs; turns the call into an edit
    r = requests.post(f"{API}/images/generations", headers=HEADERS, json=body, timeout=30)
    r.raise_for_status()
    task_id = r.json()["data"]["taskId"]

    deadline = time.time() + 300  # flat-rate id with references: p90 is about 190 s
    while time.time() < deadline:
        d = requests.get(f"{API}/images/generations", headers=HEADERS,
                         params={"task_id": task_id}, timeout=30).json()["data"]
        if d["state"] == "completed":
            return task_id, d.get("images") or [{"url": u} for u in d["resultUrls"]]
        if d["state"] == "failed":
            raise RuntimeError(f"{d.get('failCode')} retryable={d.get('retryable')}: {d.get('failMsg')}")
        time.sleep(3)
    raise TimeoutError(task_id)  # the task keeps running; look it up later, do not resubmit

task_id, images = generate(
    "gemini-3-pro-image",
    "Put this bottle on a sunlit marble kitchen counter; keep the label and the bottle shape unchanged",
    image_urls=["https://example.com/bottle.png"],
)
print(images[0])
record = requests.get(f"{API}/records/{task_id}", headers=HEADERS, timeout=30).json()["data"]
print(record["state"], record["credits"], record["currency"])

The body is model, prompt, optional image_urls for reference images, aspect_ratio (auto, 1:1, 16:9, 9:16, 4:3, 3:4, 3:2 or 2:3) and resolution (1k, 2k or 4k). Always pass resolution: on the first two ids it decides the price.

For one blocking call, send an OpenAI-style size such as 2048x1152 and no callback_url; the answer comes in the OpenAI {created, data} shape. A blocking call that runs past 95 seconds gets its headers early, so a late failure arrives as HTTP 200 with an error object in the body. With the flat-rate id and reference images that is common, so stay async there.

How long it takes

End-to-end time from submit to finished image, successful requests only, including queueing and any automatic rerouting:

Model id and workloadp50 (s)p90 (s)
gemini-3-pro-image, 1K3349
gemini-3-pro-image, 2K4983
gemini-3-pro-image, 4K85132
gemini-3-pro-image-preview, 2K5482
gemini-3-pro-image-gemini, text-to-image4590
gemini-3-pro-image-gemini, with reference images101188

The preview id at 1K and 4K and the flat-rate id at 4K had too little traffic to quote.

Three things worth acting on:

  • Reference images add time. On the default id the median goes from 39 s without references to 50 s with them, and most production calls are edits.
  • 4K is about two and a half times the 1K wait on the default id. Draft at 1K or 2K and render only the keeper at 4K.
  • A 60-second timeout is too short for 2K on every id. Poll for up to 300 seconds or use webhooks. A client-side timeout does not cancel the task: it keeps running and is billed if it succeeds, so look it up instead of resubmitting.

How often it fails, and what to retry

Over the same 7 days, content moderation refused about 2% of prompts on gemini-3-pro-image and about 4% on each of the other two ids. No request on any of the three ids failed for any other reason in that window.

That second number is mostly the gateway's doing. Each of the three ids has more than one upstream route. When a route returns busy, a rate limit or a timeout, the request moves to the next route without the caller doing anything. Of the requests that hit a route error, about 97% still came back with an image; the other 3% were then refused by content moderation. Failed tasks are not billed: the reserved amount is released and the task shows a charge of zero.

When a task does fail, the poll response carries a standard failCode and a retryable flag:

  • Retry UPSTREAM_BUSY, UPSTREAM_FAILED and TIMEOUT, at most twice, a few seconds apart.
  • Do not retry CONTENT_MODERATION, INVALID_INPUT or INSUFFICIENT_BALANCE. Change the prompt, the parameters or the balance instead.

Every task can be looked up afterwards with GET /v1/records/{task_id}, which returns the final state and the amount actually charged in USD, so reconciling a batch is one loop.

Did you get the resolution you paid for?

Completed tasks now include an images list with the real width and height of every result. Since that field was added this week, every Nano Banana Pro image we checked came back at the requested tier. For 16:9, expect 1376x768 at 1K, 2752x1536 at 2K and 5504x3072 at 4K, and compare before you store the file.

Prices, as of October 27, 2026

Price per finished image, in US dollars:

Model id1K2K4K
gemini-3-pro-image0.080.080.13
gemini-3-pro-image-preview0.0680.0680.13
gemini-3-pro-image-gemini0.030.030.03
Google standard list price0.1340.1340.24

Google bills Gemini 3 Pro Image at $120 per million output image tokens, which its pricing page converts to $0.134 per 1K or 2K image and $0.24 at 4K, plus about $0.0011 per input image. Here reference images are not billed separately and failed tasks cost nothing. Against Google's standard list price, the default id is 40% lower at 1K and 2K and 46% lower at 4K, and the preview id is 49% lower at 1K and 2K. Every parameter is on the Gemini image API reference.

A worked cost estimate

Say you are refreshing 400 product listings: three text-to-image layout drafts per product, one final 2K edit per product from the product photo, and 20 hero images at 4K for print.

Drafts   400 x 3 x $0.03 (gemini-3-pro-image-gemini) = $36.00
Finals   400 x 1 x $0.08 (gemini-3-pro-image, 2K)    = $32.00
Heroes    20 x 1 x $0.13 (gemini-3-pro-image, 4K)    =  $2.60
Total                                                  $70.60

The drafts are text-to-image, which is where the flat-rate id is closest to the default id in speed. Sending the finals to gemini-3-pro-image-preview instead brings that line to $27.20. At Google's standard list price the same 1,620 images would come to $219.20 before input-image charges.

When not to use it

  • You need Google Search grounding, for example an infographic that pulls in live facts. The request schema here covers prompt, reference images, aspect ratio and resolution; grounding is not exposed. Call Google's API directly.
  • You need masked, pixel-exact inpainting. There is no mask parameter; edits are steered by the prompt.
  • You need an image in a few seconds. Even the fastest case here, the default id at 1K, has a median of about half a minute. Nano Banana Pro is not an instant-preview model.
  • You need a direct contract with Google, for example for enterprise terms or to use committed Google Cloud spend. Use the Gemini API or Vertex AI directly.

Otherwise, run one 1K request on gemini-3-pro-image and the same prompt on gemini-3-pro-image-gemini to see whether its timing fits your queue. Each task and its exact charge show up under Records in the apimodels.app console.

Drafted with AI assistance; prices and numbers checked against our API docs and records on Oct 25, 2026.

评论

此博客中的热门博文

GPT Image 2 API: Which Model ID to Use, from $0.01 per Image

Seedance 2.0 Fast API: 480p and 720p Video from $0.071 per Second

Wan 3.0 Video API: 30-Second Clips from $0.045 per Second