GPT Image 2.5 quality auto reserves max: set quality before a batch

On Sume, quality: "auto" for GPT Image 2.5 reserves the max bound, the most expensive level, while an omitted quality uses high. So auto is not a cheap default. Before a batch, set quality explicitly, and check the endpoint pricing line with GET /v1/images/models/openai/gpt-image-2.5/endpoints.

Three behaviors to know

The docs describe how the quality field resolves, and the three cases differ in what you pay and what you reserve.

• Omitted: Sume sends high, and the estimate prices high too, so the reserve matches the run.
• auto: Sume reserves the max bound. The documented auto size also reserves the upper bound of output tokens.
• Explicit low, medium, high, xhigh or max: Sume prices exactly that level for the requested size.

What the levels cost at 1024 square

Flare and Sunburst share the same token rates: $30 per million output image tokens, $8 per million input image tokens and $5 per million input text tokens. The docs give the output estimate at 1024 by 1024 for the two top levels. Input tokens are extra, and Sume applies its list times 1.25 pricing on top.

Why this matters for a batch

A balance check runs against the reserve. If you queue 200 jobs with quality: auto, each is admitted at the max bound, so a wallet that comfortably covers 200 high images can still be short at admission. The final bill follows the completed generation, but the hold is what blocks the queue. Setting the level yourself keeps the hold close to the real cost.

A safe request

Pin the level and the size in the body. Use medium for drafts, high for finals, and keep xhigh and max for dense text or print work.

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":"Clean product photo of a steel water bottle on a stone ledge, soft morning light","quality":"medium","aspect_ratio":"4:5"}'

Check the price you will actually pay

The endpoint pricing lines are the amount Sume charges your wallet, with margin included, so you pay cost_usd times n. Read it once per model and cache it. If the line changes, your budget guard should notice. See the GPT Image 2.5 API guide for sizes and masks.