GPT Image 2.5 Sunburst Change one thing. Keep everything else.

GPT Image 2.5 Sunburst is one of the two models OpenAI shipped under GPT Image 2.5 on September 8, 2026. It is the sibling that follows edit instructions more precisely, which is why it is worth selecting when a render has to land on a specific detail. It carries the same GPT Image 2.5 quality work as Flare: better lighting, richer texture and stronger preservation of subjects from reference photos.

Both are GPT Image 2.5 and both cost the same per image. GPT Image 2.5 Sunburst is the one for precision edits on a shot that ships, and GPT Image 2.5 Flare is the faster one for drafting while the prompt is still changing. The usual split is one job, two halves: draft the set on Flare, then make the single change that matters on Sunburst.

Yes, and that is the reason to pick it. On GPT Image 2.5 Sunburst a narrow instruction lands more reliably, so a background swap, a colourway change or a label rewrite comes back closer to what you asked for. If you are generating from scratch rather than editing, the gap between the two models is small and Flare's speed is worth more.

Mostly, and not completely. GPT Image 2.5 Sunburst holds the subject, the composition and the label artwork better than the previous generation, but the edit is not pixel-locked: reflections and fine texture you did not ask to change can still move. That is why the walkthrough above ends with a side-by-side check against the original rather than a blind download.

A GPT Image 2.5 Sunburst image costs 3 to 21 credits depending on output size and quality, the same range Flare sits in, and the first render runs without a payment method. A chain of edits costs one render each, so plan for the passes rather than the finished file. Every model's rate is listed on the pricing page.

Yes, and that is the trade. GPT Image 2.5 Sunburst spends more time on each request because it is following the edit instruction more closely, so it takes longer to come back than Flare does. On a drafting loop that difference is annoying; on the one frame that goes to a client, it is the reason you picked it.

You can chain edits on GPT Image 2.5 Sunburst, and the subject holds up better than on the previous generation, but drift accumulates across a long chain. Two habits keep it under control: make one change per pass, and keep the earlier render as your reference so you can compare step four against step one. The four-step chain above is a real GPT Image 2.5 Sunburst run, not an illustration.

Yes, that is what the reference images are for. Attach the packshot along with the colourway or the model you want held, and GPT Image 2.5 Sunburst carries that look into the new frame instead of inventing a new one. Keep the reference set identical between renders; swapping one reference out mid-set is the usual reason a series suddenly stops matching.