GPT Image 2.5 quality auto: 26.35 cents reserved vs 6.59 at high

If you send GPT Image 2.5 requests with quality: "auto", Sume's pricing code estimates them at the maximum-quality price, $0.2635 for a 1024 square image, which is four times the $0.0659 you pay at high and 36 times the $0.0074 at low. Pin quality to low, medium or high and the hold equals the price you expected.

The reason is in the estimator's header: GPT Image 2.5 output is billed in image tokens, so quality sets the token grid (16, 24, 48, 64 or 96) and auto cannot know which grid the model will use. The code resolves auto to the top tier (max, a 96 grid) for the reserve. That is a conservative hold, not necessarily the final charge, and it can tie up balance on large batches.

Price by quality and size

The arithmetic

Token counts come from OpenAI's published image token calculator as implemented in gpt-image-25.ts. For a square image at the high grid of 48, the output is 48 x 48 x (2,000,000 + 1,024 x 1,024) / 4,000,000 tokens before rounding. Sume converts tokens at the fal output rate and multiplies by 1.25, which lands at $0.0659 for square high.

Because the estimate is token-based, the 16:9 row is more than three times the square row at every quality: a wider canvas has more pixels and more tokens. At high, a landscape image is $0.2225 against $0.0659 for a square one.

Steps

• Set quality explicitly on every call; do not rely on the default for batch work.
• Use low to test a prompt, medium for review images, high for finals.
• Choose the aspect ratio last: a landscape final costs about 3.4 times the square one.
• Read usage.cost after each job settles and compare it with the estimate for your quality and size.

What a batch looks like

Take 100 square images. At high the estimate is 100 x $0.065875 = $6.59. At auto the estimate is 100 x $0.2635 = $26.35, so the difference in balance you must hold is $19.76. At medium the batch is $1.65 and at low $0.74.

For a 16:9 batch of 100 the numbers are $22.25 at high and $88.95 at auto. If your workspace balance is small, a few auto jobs can look unaffordable at submit time even though the images you wanted were medium. The fix costs nothing: send the quality you mean.

Multiply any of these by 1.055 for the platform fee from Sume's pricing policy; the $6.59 high batch becomes $6.95.

What Sume does not do

Sume does not make auto cheaper than its upper bound at reserve time, and it does not claim the settled charge will equal the hold. We computed the figures above from the repo's estimator, which has an auto rule that picks the largest tier; check the final bill on your own jobs. Read on 2026-10-08.