How to find the cheapest image generation API

The cheapest API is the cheapest accepted asset

Direct answer: for predictable 1K assets, Nano Banana 2 (gemini-3.1-flash-image) is usually the cheapest starting point because its image-output leg is fixed by size — 1,120 tokens for 1K, $0.0672 official, $0.0336 for a regular B2C account after the 50% discount. For workflows already built on the OpenAI Images client, GPT Image 2 (gpt-image-2) can end up cheaper despite variable token billing, because higher acceptance and zero migration work beat a lower token rate. Anything beyond that requires a benchmark on your own briefs — and this guide shows how to run one honestly.

Run the arithmetic before choosing. One thousand accepted 1K assets on Nano Banana 2 at an 80% acceptance rate means 1,250 paid generations: 1,250 × $0.0336 = $42 of image-output spend, plus text input at $0.50/M official and any grounding reported in terminal usage. The same brief on GPT Image 2 bills image output at $30/M official ($15/M after the discount), text input at $5/M and reference images at $8/M — if it passes more often, the variable bill still wins. Acceptance rate, not the price list, decides.

See also: How image generation API pricing actually works

Match the model to the workload's cost driver

The 50% B2C discount applies to both models after exact official usage is calculated, so it never breaks a tie by itself. The gap comes from usage shape, retries and how much client code you must write and maintain for each protocol.

Price the real request shapes, not the landing page

Both models are reachable with the same prepaid key, but they speak different protocols, and the protocol decides what you can control. Nano Banana 2 uses the native Gemini generateContent shape with the x-goog-api-key header: you set responseModalities to TEXT and IMAGE, pick imageSize 1K/2K/4K and an aspectRatio, and receive the picture as a base64 inlineData part. GPT Image 2 uses the OpenAI Images routes with Authorization: Bearer: the published controls are background opaque or transparent, quality low and size as a proportion preset, the response is one non-streaming base64 PNG, and edits on /v1/images/edits accept one to five PNG, JPEG or WebP references.

# Nano Banana 2 — Gemini-native route
curl https://router.apitoken.sale/v1beta/models/gemini-3.1-flash-image:generateContent \
  -H "x-goog-api-key: $APITOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Create a square product illustration"}]}],"generationConfig":{"responseModalities":["TEXT","IMAGE"],"imageConfig":{"imageSize":"1K","aspectRatio":"1:1"}}}'

# GPT Image 2 — OpenAI Images route
curl https://router.apitoken.sale/v1/images/generations \
  -H "Authorization: Bearer $APITOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-image-2","prompt":"A clean studio product photograph","background":"opaque","quality":"low","size":"auto"}'

Do not estimate the bill from the PNG file size or pixel dimensions. For GPT Image 2 the billing authority is the terminal usage event; for Nano Banana 2 the fixed image-output leg must still be added to input, optional text or thinking output and grounding.

Run a fair per-asset cost benchmark

1. 01Pick 20–50 representative briefs from the real workload and write pass/fail criteria before any generation runs.
2. 02Give both routes equivalent inputs: same resolution class, same references, same maximum number of attempts per brief.
3. 03For every paid attempt record the request ID, terminal usage, latency and your pass/fail verdict next to the returned asset.
4. 04Divide total settled charge — including rejected outputs — by accepted assets, and compare the distributions, not the single best image.
5. 05Repeat the benchmark whenever prompts, target resolution, reference count or the provider catalog changes materially.

Savings tactics that quietly raise the bill

• Generating at 4K and downscaling every accepted asset to 1K: you paid the 4K image leg ($0.1512 official) for a 1K deliverable ($0.0672 official).
• Counting only successful outputs while hiding paid rejects, moderation failures and timed-out retries from the cost per accepted asset.
• Burning image-model tokens on text-only planning or prompt rewriting that a cheaper text model handles before the image call.
• Retrying automatically after a delivered response, or retrying without a fixed total attempt budget per brief.
• Publishing one permanent 'cheapest' verdict while live catalog availability and per-workload quality keep moving.

Validate the winner with the $5 bonus before you top up

You do not need to spend money to find your cheapest route. New accounts created with Google or GitHub receive a one-time $5 platform bonus credit — at $0.0336 per 1K Nano Banana 2 image-output leg, that covers the entire benchmark above with room to spare. Accounts registered with email and password are fully usable but not eligible for the grant. The bonus is always spent before paid balance, so the evaluation never touches your top-up.

• Top up in whole-dollar amounts by bank card or cryptocurrency; the prepaid balance never expires.
• Give the benchmark its own API key with a lifetime spending limit so a retry loop cannot drain the account.
• Keep the winning configuration pinned: model, size, references and maximum attempts — and re-test after catalog changes.

Prepaid plans: any whole-dollar amount, balance never expires →

Frequently asked questions

Related guides