GPT Image 2.5 at 1440p and 4K: output estimate by quality

At 3840x2160, GPT Image 2.5 output is estimated at $0.01113 per image on low and $0.40026 on max, before input tokens and Sume pricing. At 2560x1440 it ranges from $0.00615 to $0.2211. The 3840x2160 endpoints match Fal's GPT Image 2.5 Flare page, which lists $0.40026 for 3840x2160 at max quality; the rest come from the same token formula documented in Sume's Image API.

The numbers

Output is billed per image token at $30 per million. The token count depends on the pixel size and the quality tier. Because the quality tier multiplies the token count so strongly, going from high to max at 4K costs four times as much. Going from max at 1440p to max at 4K costs less than double.

Read it as a ladder

Two things stand out. First, the pixel count is a smaller lever than the quality tier: at low, 4K is about 1.9 times the 1024x1024 figure, while max is about 36 times low at the same size. Second, 1440p is nearly free next to 1024x1024: it adds under 10% at every tier.

So if you want more pixels for a hero image, 2560x1440 is the cheap step. Move to 4K when the layout needs it, and spend on the quality tier only where detail shows.

Constraints before you ask for 4K

OpenAI's image generation guide says custom sizes need both edges to be multiples of 16, a maximum edge of 3840, an aspect ratio between 1:3 and 3:1, and 655,360 to 8,294,400 total pixels. 3840x2160 has 8,294,400 pixels, exactly the ceiling, but 2160 is a multiple of 16 (135 x 16), so it passes.

The Sume docs repeat those rules for image_size, and add that auto quality reserves max in Sume's estimate and that named presets without a verified GPT pixel mapping reserve the output upper bound. Set an explicit quality and an explicit size when you want a predictable figure.

• Edge limit: 3840. Total pixels: 655,360 to 8,294,400.
• Both edges multiples of 16; aspect ratio at most 3:1.
• Large, high-quality requests are the most likely to return a 202 job instead of an image body.

What the table leaves out

These are output-token figures only. Text prompt tokens ($5 per million) and input image tokens ($8 per million) add to the bill when you send references, and Sume applies its own pricing on top. The authoritative figure for your account is the pricing line on the model's endpoint record, returned by GET /v1/images/models/{model_id}/endpoints.

If a 4K request does not return inside the 30-second blocking wait, Sume returns 202 with a job envelope and you fetch the image from the job result. See Jobs and results.