GPT Image 2.5 quality auto holds the max price: a 12-image batch

If you leave GPT Image 2.5 on quality: auto, Sume reserves the price of max quality. The image docs give the 1024 x 1024 output cost for max as $0.21072 before input tokens and Sume pricing, which is $0.2634 after the 1.25 factor, so twelve images hold at least 12 x $0.2634 = $3.1608. The same twelve at a fixed low quality at 1K are 12 x $0.02475 = $0.297 on the Sume list.

A hold is not a charge. A completed job captures the actual cost, and the hold in excess is released, so auto is a balance issue before it is a billing issue. A small wallet that could afford twelve low-quality images can still fail a twelve-image auto batch with a 402.

What the docs say

Flare and Sunburst use fal token rates of $30 per million output image tokens, $8 per million input image tokens and $5 per million input text tokens. Output estimates follow OpenAI's size and quality calculator. auto quality reserves max, and auto size, and named presets without a verified pixel mapping, reserve the upper bound of output tokens. Input token counts are estimates, and fal rounds the total up to $0.0001.

Pin the quality

The cheap fix is to send an explicit quality. Sume lists low, medium, high, xhigh and max; if you omit quality entirely the default is high, so auto is something you have to choose. Sume's list prices by tier at default size are $0.02475 (low, 1K), $0.055625 (medium, 2K) and $0.2225 (high, 4K), and the tier and size together set the figure.

For a catalog of twelve product shots at medium and 2K, the list price is 12 x $0.055625 = $0.6675, about a fifth of the auto hold.

When auto is the right choice

auto is for prompts where you cannot say in advance how much detail the image needs and are willing to pay for the ceiling. If the balance is large and the images are few, the hold is not a problem. In a batch, the hold multiplies by the number of accepted jobs, so check the product against the balance and the plan's accepted job capacity before you submit.

Reading the hold on the ledger

After you submit, the usage ledger shows a reserved row at the estimated amount. When the job completes, the captured row is the actual cost and the unused part of the hold is released. Compare the two on the first image of a batch: if the captured amount is far below the reserved one, pin a lower quality for the rest and recover headroom. The numbers in this post cover output tokens at one size only; input tokens for references and prompts are extra, and fal rounds the total up to $0.0001.