How many GPT Image 2.5 images does $10 buy on Sume?

Ten dollars buys 1,355 GPT Image 2.5 images at low quality, 606 at medium, 151 at high, 85 at xhigh and 37 at max, all at 1024x1024 on Sume. The counts come from dividing the budget by Sume's quoted price per image and rounding down. Change the quality and the count moves by a factor of up to 36.

Both ChatGPT Image 2.5 ids, openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst, quote the same amounts. Sume bills only completed generations; a failed one is not charged, so a failed call does not use up budget.

The table

Prices are Sume's quote at a fixed 1024x1024 size with a one-character prompt. A real prompt adds a few input text tokens at $5 per million, which is a fraction of a cent for a long paragraph, so treat each count as an upper bound.

Why the list is not the whole bill

Two details keep a real spend a little above the division. First, a job holds its estimate while it runs: if you leave quality as auto, Sume reserves the max price, so every in-flight auto job holds the larger amount even if it would settle lower. Second, Fal rounds each total up to $0.0001, which adds at most a hundredth of a cent per image.

Neither detail changes the ranking of the tiers. They only matter when you size a prepaid balance or a daily cap, so pin quality and image_size and the hold equals the price in the table.

Run it yourself

The script below reproduces the table and works for any budget. Change budget, or add a row for another size from the quote you get back.

PRICES = {  # Sume quotes, USD per 1024x1024 image
    "low": 0.007375,
    "medium": 0.0165,
    "high": 0.065875,
    "xhigh": 0.117125,
    "max": 0.2635,
}
budget = 10.00
for quality, price in PRICES.items():
    count = int(budget // price)
    print(f"{quality:>6}: {count:5d} images, ${count * price:.2f} spent")

What moves the count

Four inputs change how far a budget goes. All four are documented on the Image API page.

• Size. A 1536x1024 or 1024x1536 image quotes lower than a square one at every tier, because output tokens scale with the pixel grid. A 3840x2160 image quotes higher.
• References. An edit adds input image tokens at $8 per million. Sume's quote for a one-reference edit at medium is above the text-to-image quote for the same size.
• n. Each extra image is another full charge: the cost is the per-image price times n.
• Omitted settings. Leaving out quality means high, and leaving out the size reserves the upper bound of output tokens while the job runs.

A sensible way to spend the ten dollars

A budget split is easier to defend than a single tier. Spend $2 on 271 low drafts, then $8 on 121 high finals, and you have 271 candidates to choose from before you pay for a final. Spend the same $10 on max alone and you have 37 images with nothing to choose between.

If the images feed a larger job such as a video first frame, check the aspect ratio before you scale up, since the same budget buys fewer frames at 4K.

One more habit helps: write the per-image price next to the budget in whatever tracker you use, and update it from usage.cost after the first real call. A price you read from a table last month is a guess; a price you read from your own response is a fact. The table here was computed on 2026-10-05, and the model's rates are set by Fal, so a later change in the upstream rate would change every count above. For the monthly view of the same arithmetic, see 10,000 images a month.