GPT Image 2.5 quality auto is estimated as max: set a tier on Sume

Sume's estimator treats quality auto on GPT Image 2.5 as max, up to $0.21 per 1024x1024 image in output tokens. When auto is fine and when to pin a tier.

4 min readSume
All posts

Sending quality: "auto" to GPT Image 2.5 on Sume is not a cheap default. The Image API docs say auto quality reserves max, which is estimated at $0.21072 in output tokens for a 1024x1024 image, against $0.05268 for high. If you want the model's own choice, you can have it; if you want a predictable bill, set low, medium, high, xhigh or max yourself.

What the docs say

The Image API page states that for Flare and Sunburst auto quality reserves max, and that auto size, and named presets without a verified GPT-specific pixel mapping, reserve the output token upper bound. It also says omitted quality defaults to high. So three spellings give three different estimates.

OpenAI's image generation guide says both models default to auto, and lists low, medium, high, xhigh, max and auto as the values.

Estimated output per 1024x1024 image by quality input, before input tokens and Sume pricing (read 2026-10-04)
You sendSume treats it asOutput estimate
(omitted)high$0.05268
"high"high$0.05268
"auto"max (reserve)$0.21072
"max"max$0.21072
"low"low$0.00588

Reserve is not the same as the final charge

The docs use the word reserve for the estimate; they do not promise that the final charge equals it. What you can rely on is the rule on the Billing and cancellation section of the same page: a completed generation is billed per endpoint pricing, and a failed one is not billed. Test your own prompt with n: 1 and read usage.cost.

Plan around the upper figure when a wallet balance is small. A request estimated at $0.21 against a $0.10 balance is a request you may not be able to afford, while the same prompt at medium costs about $0.013.

When auto is a fair choice

  • A one-off exploration where you do not care about the spend.
  • A prompt you have not tuned and do not know how much detail it needs.
  • Never for a loop, a batch, or a user-facing feature that loops on its own.

The size half of the rule

Size has the same trap. auto size reserves the output token upper bound (per the Image API page), so pair a pinned quality with a pinned size such as aspect_ratio: "1:1" or explicit pixels when you need a tight estimate. The Errors and credits page covers what happens when a balance is too low for an estimate.

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": "A red bicycle against a white wall",
    "quality": "medium",
    "aspect_ratio": "1:1"
  }'

Sources

Related posts

More in Pricing

All Pricing posts

Written by Sume