GPT Image 2.5 quality auto reserves max on Sume: pin quality
Sume's image docs say quality auto reserves max, and auto size reserves the output token upper bound. What to set instead, plus the 1024 by 1024 figures.

On Sume's GPT Image 2.5 models, quality: auto reserves the max price, and auto size reserves the output token upper bound. If you know the tier you need, name it, and the amount held before the job runs gets smaller.
OpenAI's guide lists the quality values as low, medium, high, xhigh, max and auto. Sume defaults to high when the field is omitted.
What the docs state
Figures from Sume's Image API page, read 2026-10-03. Sume prices are before Sume's own pricing is applied, as the page says.
| Item | Value |
|---|---|
| Output image tokens | $30 per million |
| Input image tokens | $8 per million |
| Input text tokens | $5 per million |
| 1024 by 1024, xhigh output | $0.09366 before input tokens and Sume pricing |
| 1024 by 1024, max output | $0.21072 before input tokens and Sume pricing |
quality: auto | Reserves max |
auto size, or presets without a verified pixel mapping | Reserve the output token upper bound |
Omitted quality | Defaults to high |
What the hold means
Before a paid job runs, the platform reserves funds against its worst case. Sume's errors and credits docs cover what happens when a wallet cannot cover a request, so a large reservation matters most for small balances or many parallel jobs. The docs state the reservation rule for auto; they do not publish a hold amount for a given request, so read the live cost from the response rather than computing it here.
A request that pins quality
Setting quality removes the max reservation. For size, the docs allow custom pixels through image_size: both edges multiples of 16, a maximum edge of 3840, an aspect ratio of at most 3:1, and 655,360 to 8,294,400 pixels. The docs excerpt used here does not show the value's JSON shape, so read it from the model's catalog descriptor before sending one. The request below pins quality only.
import os, requests
body = {
"model": "openai/gpt-image-2.5",
"prompt": "a ceramic mug on a wooden desk, morning light",
"quality": "medium",
}
r = requests.post("https://api.sume.com/v1/images",
headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]},
json=body, timeout=60)
print(r.status_code, r.json())Choosing a tier
A starting rule, not a benchmark.
- Drafts:
lowormedium. You are judging composition. - Finals:
high, the default, unless a close look shows a flaw that a higher tier fixes. xhighandmax: only after a side by side shows a visible gain on your content. The 1024 by 1024 figures above show max output costing more than twice xhigh.- Leave
autofor cases where you want the model to decide and your balance can carry the reservation.
Sources
Related posts
More in Pricing
- Grok Imagine video: cost of a 10-second clip on each tier
xAI lists grok-imagine-video-1.5 at $0.080 per second, 1.5-lite at $0.020 and the base model at $0.050. The 10-second math and a draft-then-final plan.
- Is the Veo 3.1 API free? The per-second prices Google lists
Veo 3.1 on the Gemini API is priced per second for Standard, Fast and Lite, charged only on success. The rates, plus 8-second clip totals and what to verify.
- LTX 2.5 Pro at $0.17 a second: 10-second cost and open weights
Higgsfield lists LTX 2.5 Pro at $0.17 per second, so 10 seconds is $1.70. How that compares with running LTX locally and what Sume's catalog lists.
- Luma credits run 4 to 314 per second: a 10 s clip costs $0.12 to $9.42
Luma's pricing page lists video at 4 to 314+ credits per second. On the Plus plan that is $0.12 to $9.42 for one 10 second clip. The math and a budgeting habit.
Written by Sume