Nano Banana API Python Example: requests, No SDK

Most "Nano Banana API" write-ups stop at a curl command. This one is Python: a working requests script that generates an image, edits it, and reads back the result — no Google Cloud SDK, no provider-specific client library, just requests and one API key.

Set up the client

Nano Banana (Google's Gemini image model) sits behind the google_banana provider on the gateway. You only need requests and your API key — no google-generativeai package, no service account JSON:

import os
import uuid
import requests

API_KEY = os.environ["AIGENWAY_API_KEY"]  # sk_live_...
BASE_URL = "https://api.aigenway.com/v1"

session = requests.Session()
session.headers.update({
    "Authorization": f"Bearer {API_KEY}",
    "Content-Type": "application/json",
})

Confirm the exact model codes live on your account before hardcoding one — providers add and retire versions on their own schedule:

resp = session.get(f"{BASE_URL}/methods", params={"provider": "google_banana"})
resp.raise_for_status()
print(resp.json())  # model codes, supported methods, current pricing

Generate an image from text

POST /v1/generations with provider, model, method, params, and an Idempotency-Key so a retried request never double-charges you:

def create_generation(provider, model, method, params):
    resp = session.post(
        f"{BASE_URL}/generations",
        headers={"Idempotency-Key": str(uuid.uuid4())},
        json={
            "provider": provider,
            "model": model,
            "method": method,
            "params": params,
        },
    )
    resp.raise_for_status()
    return resp.json()

task = create_generation(
    "google_banana",
    "gemini-3.1-flash-image-preview",
    "text_to_image",
    {
        "prompt": "a cinematic neon-lit ramen stall in the rain, shallow depth of field",
        "aspect_ratio": "16:9",
    },
)
print(task["id"], task["status"])

Edit an existing image

image_edit takes one source image plus an instruction — pass the source as a URL in input_images and reuse the same create_generation helper:

edit_task = create_generation(
    "google_banana",
    "gemini-3.1-flash-image-preview",
    "image_edit",
    {
        "prompt": "replace the background with a snowy mountain range at golden hour",
        "input_images": ["https://example.com/portrait.jpg"],
    },
)
print(edit_task["id"], edit_task["status"])

Keep edits to a single source image per request, and be explicit about what should stay the same ("same pose, same lighting") — vague prompts produce inconsistent edits.

Poll or webhook for the result

Both calls above return a task, not a finished image — generation happens async. Poll it in a small loop, or skip the loop entirely with a webhook.

import time

def wait_for_result(task_id, timeout=120, interval=2):
    deadline = time.time() + timeout
    while time.time() < deadline:
        resp = session.get(f"{BASE_URL}/generations/{task_id}")
        resp.raise_for_status()
        data = resp.json()
        if data["status"] in ("succeeded", "failed"):
            return data
        time.sleep(interval)
    raise TimeoutError(f"task {task_id} did not finish in time")

result = wait_for_result(task["id"])
print(result["status"], result.get("outputs"))

For production code, skip the polling loop and set callback_url on the original request instead — the gateway POSTs the finished task to your endpoint the moment it's ready, signed so you can verify it came from Aigenway. The task shape, status values, and webhook signature are identical for every model on the gateway, so this same wait_for_result helper works unchanged once you add Imagen or a video model. Full contract in the task lifecycle docs.

Why this beats a provider SDK

• No SDK dependency. requests (or any HTTP client) is enough — nothing to pin, nothing to upgrade when Google ships a breaking client release.
• Same script, every model. Swap provider/model and the identical create_generation/wait_for_result functions call GPT Image, Imagen 4, or Kling — you write the async handling once.
• Prepaid USD billing. No Google Cloud billing account, no per-token math — check your balance from the API.
• Idempotent by default. The Idempotency-Key header means a retried request on a flaky connection never generates (or bills) twice.