What Drives Your Image Generation Bill

Most people think resolution or prompt length drives image API costs. They are wrong.

We tested image generation across four variables: model, resolution, image count, and quality. Here is what we found.

The Quality Trap

The biggest cost lever is the quality setting. Most APIs let you choose low, medium, or high quality.

For models like gpt-image, you pay per token. High quality does not mean more pixels. It means more rendering effort.

We tested the same prompt at the same resolution. The file size stayed the same. However, the high-quality setting cost 36x more than the low-quality setting.

You pay for the work the model does, not the size of the file.

Key Findings

• The image is the cost, not the prompt. In text-to-image, your prompt is a rounding error. If you use image-to-image, the input image becomes a massive cost.

• Model choice is a 6x lever. Different billing shapes change your total spend. Some models charge per token. Others charge a flat fee per image.

• Resolution is a weak lever. Increasing resolution does not always increase cost proportionally. In some models, the cost stays flat regardless of size.

• Caching does not work. Many people try to use prompt caching to save money. Image generation is stateless. You cannot reuse prompt state to lower your bill.

• Batching does not save money. Generating four images in one call often bills you for the prompt four times. There is no cost sharing.

How to Pick a Model

If you need low-quality drafts or thumbnails: Use a per-token model. It is the cheapest option for low-fidelity work.

If you need medium or high-quality images: Use a flat per-image model. Per-token costs explode at high quality. A flat fee is more predictable and usually cheaper for professional work.

If you are doing image-to-image editing: Default to flat per-image models. They often include reference images for free. Per-token models charge extra for every reference image you provide.

Stop defaulting to high quality. Set your quality to the job.

Source: https://dev.to/synthorai/what-actually-drives-your-image-generation-bill-2bc2