12 images from an agent on GPT Image 2.5: 12 cents at low, 72 at high

On Sume, a 1024-class GPT Image 2.5 image costs 1 cent at low quality, 2 cents at medium and 6 to 7 cents at high. Sume's image docs say that if you omit quality, the default is high. So an agent that asks for twelve images and says nothing about quality spends 72 to 84 cents, where twelve low-quality drafts cost 12 cents. State the quality in your instruction.

What the docs say

Per the Image generation page (read 2026-10-08), ChatGPT Image 2.5 is available as openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst. Both take quality of auto, low, medium, high, xhigh or max. If you omit it, the default is high. auto quality reserves max, so it is the setting to avoid when you want a predictable bill. Custom sizes need both edges to be a multiple of 16, a maximum edge of 3840 and an aspect ratio of at most 3:1.

• Default quality when omitted: high.
• auto reserves max at submit time.
• Pixels must be between 655,360 and 8,294,400.

Twelve images, three settings

Sume bills the catalog price times 1.25, rounded up to the cent. The figures below are billable cents at the 1024-class sizes.

A draft-then-finish pattern

Ask for eight low-quality drafts first: 8 x 1 = 8 cents. Pick the two you like, then regenerate those at high: 2 x 7 = 14 cents at the upper end. The whole job costs 22 cents at most, against 8 x 7 = 56 cents if you had asked for eight at the default quality.

Spell it out in the tool call. Tell the agent to read the live contract with tools_schema for generate_image before it submits, because the MCP schema is the authority for which fields the tool accepts. Then run dry_run=true once, and give each image its own idempotency_key.

Checks that save money

Keep the instruction explicit: model, quality, size and count. A prompt that reads "a few images" invites a larger batch, and Sume's max_spend_usd only protects you if you provide it. For twelve high-quality images, 0.84 is a sensible cap; for twelve low, 0.12. After the wave, usage_get shows the debited amounts, which are the figures to quote to the user.

Why high is the default that surprises

Most people assume an unspecified quality means a middle setting. Here it means high, the second most expensive of the four common tiers, and the same instruction without a quality word can cost six or seven times what a low draft does. Sume also documents xhigh and max; at 1024x1024, the underlying output estimate is $0.09366 for xhigh and $0.21072 for max before input tokens and Sume's pricing, so those two are not suited to bulk runs.

If you let the agent choose auto, remember that Sume reserves max at submit time. The charge settles on the real result, but the reservation can block a call when the balance is low. Pass an explicit tier and the reservation matches what you expect.