Nano Banana API: price per image, request and code

What a Nano Banana 2.1 or Nano Banana Pro image costs

Credits per run from a prompt, for one image and for 4, and for one image with one reference image. Dollars are at the headline credit price; plans with more credits pay less per credit.

Computed from the quote the API charges with, as of 15 Nov 2026. The price is the same for every aspect ratio.

One request, then read the job

POST /api/v1/images charges the run and answers with a 201 job in queued or running. Read GET /api/v1/jobs/{id}?wait=50: it holds the line for up to 50 seconds and returns the moment the job is final, with the URLs in output.imageUrls. The body below is the cheapest valid Nano Banana 2.1 request, one square image, which costs 8 credits.

curl -X POST https://kinoseed.app/api/v1/images \
  -H "Authorization: Bearer ks_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a ceramic mug on a sunlit desk, product photo",
    "model": "nano-banana-2-1",
    "aspect": "1:1",
    "count": 1
  }'

Send the same body to Quote first if you want the number before you spend anything; it charges nothing. Fields and errors are on Images and Jobs.

The same run from an agent over MCP

An agent connected to the MCP server sends the same arguments to the generate_image tool, which waits for the run and returns the image URLs. Results are URLs, never inline base64, so a batch of pictures does not fill the context window. Setup for Claude Code, Cursor and others is in the MCP server for image generation guide.

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "generate_image",
    "arguments": {
      "prompt": "a ceramic mug on a sunlit desk, product photo",
      "model": "nano-banana-2-1",
      "aspect": "1:1",
      "count": 1
    }
  }
}

What Nano Banana 2.1 or Nano Banana Pro accepts

Every field is optional except prompt. Without model the run uses the catalog default, so send the id you want.

• aspect: a ratio such as 16:9 or 9:16; 1:1 when it is left out.
• count: 1 to 8 images in one run; the price multiplies.
• referenceImageUrls: up to eight pictures to edit or follow, as public https URLs or URLs from POST /files.
• seed: a whole number, to repeat a result.
• metadata: your own ids, returned on the job untouched.

On Nano Banana 2.1 a run with a reference image goes to the model's edit endpoint and costs 9 credits instead of 8.

On Nano Banana Pro a run with a reference image goes to the model's edit endpoint and costs the same.

The API refuses private addresses, localhost and plain http in referenceImageUrls before anything fetches them or anything is charged.

What moves the price

The model and the number of images. The aspect ratio does not change it. 4 images on Nano Banana 2.1 in one run are 30 credits; 4 images on Nano Banana Pro in one run are 110 credits.

For drafts, a cheaper model often does the job: one FLUX 3 Klein image is 3 credits. Test the prompt there, then run the keeper on Nano Banana 2.1 or Nano Banana Pro. Every image model's price is side by side in Image generation API with credits.

When to pick Nano Banana 2.1 or Nano Banana Pro

Pick Nano Banana 2.1, Google's newest image model, for editing a photo you add: send it in referenceImageUrls and describe the change. Nano Banana Pro costs more per image; weigh it against 2.1 on the same prompt. Studio users get both in the Image studio; the Nano Banana 2.1 and Nano Banana Pro model pages have their prices and specs.

Nano Banana 2 (nano-banana) runs too and is in the price table. For posters and pictures with a lot of text, compare with GPT Image 2 in Nano Banana Pro vs GPT Image 2.

Side by side with the models people weigh it against: Nano Banana 2.1 vs Nano Banana 2, Nano Banana 2.1 vs Nano Banana Pro, FLUX.2 vs Nano Banana Pro, Nano Banana 2 vs Nano Banana Pro, GPT Image 2.5 vs Nano Banana Pro, FLUX 3 vs Nano Banana Pro, Seedream 5.0 · Lite vs Nano Banana Pro.

Failures, retries and limits

A run that fails is refunded automatically, and the job says so, so a provider error never costs credits. Identical requests sent close together return the same job instead of charging twice, and an Idempotency-Key header makes that explicit for retries from your own code.

An account runs up to 5 generations at once; one more gets a 429 too_many_running error with a Retry-After header, before anything is charged. Results carry no watermark. Rate limits and error codes are on Errors and Rate limits.