Which Sume image models accept quality? Only five do

In the image catalog I read in Sume's repository on 2026-10-03, five rows list a quality field: openai/gpt-image-2, openai/gpt-image-2.5, openai/gpt-image-2.5-sunburst, ideogram/ideogram-v3, and ideogram/ideogram-v4.5. The other fourteen rows do not, and sending quality to one of them returns 400 unsupported_parameter. So a shared request builder that always sets quality: "high" will break on Nano Banana, Seedream, Flux, Grok, and the rest.

The Image API page gives the rule: a request that sets a parameter the selected model does not list is rejected rather than dropped. The values come from the catalog descriptors, which you can read yourself at GET /v1/images/models.

What values does each of the five take?

The GPT Image 2.5 pair has the longest list. The docs state quality: auto|low|medium|high|xhigh|max and say an omitted quality defaults to high. GPT Image 2 stops at high. Ideogram 4.5 takes low, medium, and high, and its omitted value is medium.

What happens when I send xhigh to the wrong model?

The code depends on where the mismatch is. xhigh to GPT Image 2 is invalid_request with supported: ["low", "medium", "high"], because the model has the field but not that value. Any quality to Nano Banana 2 is unsupported_parameter, because the model has no such field. Both are 400s, and neither bills a generation.

The old route differs: Image 1.0 lists quality as low (default), medium, or high, and says to escalate for finals, dense text, or packaging. It is a compatibility alias for Auto, so use the Images API for new work.

How do I write one builder that works across models?

Fetch the catalog and add quality only when the descriptor exists and your value is in it. When it is absent, move the quality demand into the prompt for text-heavy work, or choose a model that has the field. A GPT Image 2.5 draft at low and a final at high is one pattern on that model; on a model with no field there is nothing to escalate, so pick the model by the job instead.

What does quality cost on the models that have it?

For the GPT Image 2.5 pair, the docs say auto quality reserves max, so an auto request holds the largest amount before it settles. At 1024×1024 the page quotes xhigh output at $0.09366 and max at $0.21072 before input tokens and Sume pricing. Ideogram 4.5's list prices by quality are $0.03, $0.06, and $0.22. Read the endpoint pricing line for the amount Sume charges.

Because the tiers are priced differently, the safest default for a shared builder is to leave quality unset and let each model apply its documented default, then raise it only for the images that need it.