10,000 images on Sume: $100 on Soul up to $2,700 on ChatGPT Image 2

Ten thousand images on Sume cost between $100 (Higgsfield Soul, 1 cent each) and $2,700 (ChatGPT Image 2, 27 cents each) at the high-quality list rate. Sume bills the catalog list price multiplied by 1.25 and rounded up to the next cent. The middle of the catalog, such as Flux 2 Pro at 4 cents or Seedream 5.0 Lite at 5 cents, lands at $400 to $500 per 10,000.

Cost grid for every catalog model

Each row is the catalog list price per image multiplied by 1.25 and rounded up to a whole cent, then multiplied by the image count. The GPT rows use the high-quality 1024 display rate, and Nano Banana rows use the 1K default, as the catalog notes say. Larger sizes or other qualities change the charge for those models.

How to read the grid

The spread is 27 to 1. Soul bills 1 cent per image and ChatGPT Image 2 bills 27 cents, so 10,000 images differ by $2,600. ChatGPT Image 2.5 sits at 7 cents for high quality, about a quarter of ChatGPT Image 2, which is why its row is much closer to the cheap end.

Failed or cancelled generations are not billed, so these totals count only completed images. Sume bills each completed image in full and a failed one not at all.

Ten thousand images is also a good moment to stop sending them one by one. Most catalog models take n up to 4 per call, so 10,000 images is 2,500 calls. Grok Imagine takes 1 per call, so it needs 10,000 calls.

• Cheapest: Higgsfield Soul at 1 cent, then Grok Imagine, Qwen Image and Imagen 4 Fast at 3 cents.
• Most expensive: ChatGPT Image 2 at 27 cents, Nano Banana Pro at 19 cents.
• Same price on both rows: ChatGPT Image 2.5 and its Sunburst twin.

Before you commit to a number

Run 100 images first. The 100-image column is a cheap test of the model you picked, and it tells you the quality tier and ratio you actually need. Read the live price for any model with GET /v1/images/models/{model_id}/endpoints, because the endpoint pricing line is the amount charged to your wallet and already includes the margin.

Running a batch this size

Slow settings can degrade a synchronous call to a 202 job envelope. POST /v1/images waits up to 30 seconds and returns 200 with the image, or 202 with a job. Read the status code, not the body shape, then poll GET /v1/jobs/{id}/status and fetch GET /v1/jobs/{id}/result. For a long batch, send mode: async, or mode: webhook with a public HTTPS webhook_url, so no client sits waiting.

Billing is all or nothing: a completed generation bills in full, a failed or cancelled one does not, and a client disconnect counts as a failed generation. That means a retry loop does not double-charge a failed attempt, but it will pay for every success, so de-duplicate prompts before you queue 10,000 of them.