1,000 GPT Image 2.5 images: $10 on low up to $270 on max, on Sume

A thousand GPT Image 2.5 images at 1024-class size cost about $10 at low, $20 at medium, $70 at high, $120 at xhigh and $270 at max on Sume. Quality alone moves the price 27 times, so choose the tier before you choose the count. If you omit quality, the default is high.

Per-tier cost

Low, medium and high come from the Sume billed rates for 1024-class sizes: 1, 2 and 7 cents (the catalog list for high is $0.0527, times 1.25 is $0.0659, rounded up to 7 cents). xhigh and max come from the Image API docs: at 1024x1024 the output estimates are $0.09366 and $0.21072 before input tokens and Sume pricing. Multiplying by 1.25 gives $0.1171 and $0.2634, which round up to 12 and 27 cents. Treat those two as estimates, since input tokens are added at admission.

What the tiers mean for a plan

The ladder is steep at the top. High to xhigh adds 5 cents per image, and high to max adds 20. For 1,000 images that is $50 and $200 extra. The docs also note that quality auto reserves max, so an auto request holds the max amount while it runs. Pin a tier when you care about the bill.

Both gpt-image-2.5 and gpt-image-2.5-sunburst use the same Fal token rates and the same limits, so the numbers above apply to either row. Both support up to 16 image references, an optional mask_url and background: auto, transparent or opaque.

• Draft sweeps: low at 1 cent, so 1,000 candidates is $10.
• Production batches: medium at 2 cents is $20 per thousand.
• Final art: high at 7 cents, $70 per thousand; xhigh and max only for hero pieces.

A simple rule

Generate wide at low, pick winners, then re-run only the winners at high or above with the same prompt. A hundred winners at max is $27, which is cheaper than generating a thousand at xhigh by more than a factor of 4.

Size and the 1024-class caveat

All of the figures above are 1024-class sizes. For custom pixels, both edges must be multiples of 16, the maximum edge is 3840, the aspect ratio must be at most 3:1, and the image must have between 655,360 and 8,294,400 pixels. Named presets and auto size without a verified pixel mapping reserve the upper bound of output tokens, so a large or auto size can reserve more than the figures shown. If the bill must stay predictable, pin the size and the quality in every request.