Cheapest image on Sume: GPT Image 2.5 low $0.02475 vs grok $0.025

The cheapest image row on Sume's price list is GPT Image 2.5 at low quality and 1K, at $0.02475 per image. That is 0.025 cents under grok-image and imagen-4-fast, which both list at $0.025. Over 10,000 images the gap is $2.50. If your only criterion is the lowest billed price per image, GPT Image 2.5 low is the answer today, with the condition that you ask for low quality and a 1K-class size.

That condition matters. GPT Image 2.5 is the only row in this group that Sume prices by quality and size, so the $0.02475 figure does not hold if you send the model's default. The Image API docs say that omitting quality on GPT Image 2.5 means high.

Per-image and per-batch arithmetic

All prices are the provider list price times 1.25, as Sume lists them for default options. The batch columns multiply the per-image price by the batch size, with no rounding other than the cents shown.

The default-quality trap

On /v1/images, GPT Image 2.5 takes quality: auto|low|medium|high|xhigh|max, and the default when you omit it is high. The Sume list shows high at 4K as $0.2225, and the catalog default estimate is that same figure. A job that omits quality can therefore cost 8.99 times the low-quality price ($0.2225 divided by $0.02475). Teams that migrate a cheap draft job from another model often hit this first.

Auto quality is stricter still: the docs state that auto reserves max when Sume checks your balance, so it is not a cost-saving mode either.

How to pin the cheap row

Send the quality explicitly, and keep draft sizes small. The request below pins low quality on the GPT Image 2.5 model that Sume lists as openai/gpt-image-2.5.

Before you commit to a volume, run ten calls and add up usage.cost from the responses; Sume bills each completed image in full and does not bill failed ones. See the Image API docs for the billing rules.

curl -s https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-image-2.5","prompt":"flat icon of a red kettle","quality":"low"}' \
  | python3 -c "import sys,json; print(json.load(sys.stdin)['usage']['cost'])"

What the quarter cent buys at scale

The gap per image is $0.00025. At 4,000 images a month it is $1, and at 100,000 images a month it is $25. That is real money only at high volume, and at that volume the quality setting matters more than the row. Compare the spread inside GPT Image 2.5: low at 1K is $0.02475, medium at 2K is $0.055625, and high at 4K is $0.2225. Moving one step up the quality ladder costs more than the whole gap between GPT Image 2.5 low and grok-image, so the first decision is always the quality step, and the row second.

Two of the cheap rows share a price. grok-image and imagen-4-fast both list at $0.025. Sume does not publish a pixel size for each row on the price list, so check the model's descriptors on GET /v1/images/models for the sizes and ratios it accepts before you move a workload.

When the cheapest row is the wrong one

A quarter cent per image rarely decides a project. The row you pick should follow the job: references, masks, text rendering and size all differ by model. Use the cheapest row for drafts, thumbnails and tests, then pay for a higher row only for the images you keep.