Flare vs Sunburst

The November 4 API ships two models. Flare is the smaller one for interactive drafts. Sunburst is the base model for detail and for edits that have to keep a product or a person consistent. OpenAI’s prompting guide describes that split directly.

Use Flare while the idea is still moving: a backdrop, a pose, a palette. The announcement says a low-quality Flare image can land in about six seconds. That is the point of the model, not a benchmark we re-ran.

Use Sunburst for the file you will publish or edit again. Type, small product parts, and identity across a second run are the cases that justify the wait. Both models can do GIF and SVG. Both accept the complexity control.

On this site the charge does not change. The label beside the button stays at 20 credits for either model. OpenAI’s token rate is the same per image-output token, and Sunburst can still consume more tokens on a dense prompt. The flat credit is so the price is visible before you click, not a claim that both calls cost OpenAI the same.

A practical loop: Flare at medium quality to pick a direction, then Sunburst at high quality on the winner. Raise complexity only when the layout is dense. Lower it for a logo or a one-object recolor.