GPT Image 2.5 low is $24.75 per 1,000: 25 cents under 2.5-cent rows

The short answer

For a 1,000-image draft run, GPT Image 2.5 at low 1K costs $24.75 on Sume (1,000 x $0.02475), while grok-image, imagen-4-fast and qwen-image each cost $25.00 (1,000 x $0.025). The gap is $0.25 per 1,000 images, which is 1%. It is a tie in practice, so choose by the parameters each row accepts.

The four rows

All four prices come from the Sume price list as of 2026-10-09. The GPT Image 2.5 price is its low 1K point, and the model also has medium and high points that cost more. Ratio lists come from the repo's image catalog.

The catch with GPT's default

The $0.02475 price is for low quality. If you leave quality out of a /v1/images request, the docs say the default is high, and auto is estimated as max. Those cost far more than low. A draft run therefore must set quality: "low" on every call. Image 1.0 behaves the other way, with low as its default, which is a trap when you migrate.

Low quality is also a quality decision. Look at ten low results before you commit to 1,000.

Making a draft run safe

Wrap the request in a small function that always sets quality, and refuse to send a body that lacks it. A missing field turns a $24.75 run into a far larger one. The catalog list shows gpt-image-2.5 at medium 2K as $0.055625 and at high 4K as $0.2225, so 1,000 images at those points cost $55.63 and $222.50. Read usage.cost on every response and stop the run if the total passes a limit you set before you began.

When the cheap row is the wrong choice

A low-quality draft is for layout and idea checks. Text, fine detail and faces are where a low tier shows its limits. If a draft passes the layout check, re-render the winner at medium or high.

At the listed rows, the move from low 1K ($0.02475) to medium 2K ($0.055625) costs 3.1 cents more per image, which is cheap for a final.

Pick GPT Image 2.5 if you need masks, background: transparent, or up to 16 references, which the Sume docs list for it. Pick Imagen 4 Fast if you want the plainest 16:9 or 9:16 output and do not need 4:5. Pick Grok Image or Qwen Image if you want the longer ratio lists. A mixed plan works too: draft 1,000 images on one row for about $25, shortlist 50, and re-render those at a higher tier.

Another check is to set a spend alert in your own code. Add up usage.cost as responses arrive and stop when the total passes the number you set, which for a 1,000-draft run at $24.75 might be $26.