GPT Image 2.5 xhigh vs max at 1024x1024: $0.09366 vs $0.21072

For a 1024x1024 image, Sume's Image API docs estimate GPT Image 2.5 xhigh output at $0.09366 and max output at $0.21072, both before input tokens and Sume pricing. Max costs about 2.25 times xhigh at that size. Use max for the one hero frame you will print or crop hard, and xhigh or lower for everything else.

Where the two numbers come from

The docs use the output rate of $30 per million output image tokens, which OpenAI's launch report also states for both API models. Output estimates follow OpenAI's size and quality calculator. The figures are provider-side estimates that exclude input tokens, so they are a floor for planning, not the amount on your invoice.

Auto reserves the top tier

The docs state that auto quality reserves max. If you leave quality as auto on a large batch, your wallet reservation is sized for the most expensive tier even if the image would have been cheaper. Pass quality explicitly. If you omit it entirely, the default is high.

When max is worth it

Use max for a hero image with small text, dense packaging or a large print. Use xhigh for storefront listings viewed on a phone. Draft at low, choose the composition, then re-run the winning prompt once at the final tier.

curl -X POST "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":"hero shot of a glass perfume bottle with embossed label text, dark gradient backdrop","quality":"max","aspect_ratio":"1:1"}'

Limits

xhigh and max are accepted only by ChatGPT Image 2.5. Other models return 400 unsupported_parameter for a quality their catalog entry does not list. Large renders can also exceed the 30-second wait and come back as a 202 job envelope, so read the status code and fetch the result from the job endpoint when that happens.

A tiering rule

One way to assign tiers across a campaign:

• Layout drafts: low.
• Social and listing assets: medium or high.
• Final hero or print: xhigh, and max only for the one frame that needs it.