GPT Image 2.5: What Changed and How to Prompt the New Flare and Sunburst Models

OpenAI shipped GPT Image 2.5 on September 8, 2026, its first major image-model update since GPT Image 2. It ships as two variants - Flare and Sunburst - available through ChatGPT, Codex and the API, and the headline number is latency: OpenAI says generation is up to 50% faster than GPT Image 2 at comparable quality.

What actually changed

• Sharper detail and more natural lighting and textures, especially on skin, fabric and reflective surfaces.
• Better preservation of the subject in a reference photo across an edit, instead of the face or product drifting a little with each pass.
• More reliable multi-turn editing - a sequence of small instructions ("now make the background darker", "move the logo left") holds up better than it did before, without earlier edits getting undone.
• Two variants instead of one size-fits-all model: Flare for speed, Sunburst for control.

Flare vs Sunburst: which one to reach for

Flare is the default for most work - product shots, thumbnails, social posts, quick concept passes - because it is faster and still a step up from GPT Image 2 on quality. Reach for Sunburst when a single image is going through several rounds of edits and needs to survive all of them without losing the details you already locked in: a hero campaign image, a product mockup that has to match brand guidelines exactly, or anything where the client will notice if the logo shifts two pixels between drafts.

What this means for your prompting habits

The biggest practical change for creators is that multi-turn editing is now dependable enough to actually use as a workflow, not just a party trick. Instead of writing one long prompt trying to describe everything at once, you can generate a rough pass, then correct it in short, specific instructions - one change per turn - and expect the model to keep what it already got right.

The same discipline pays off on any image model, including the ones already in Karya: be specific about what should change and say nothing about what should stay the same (the model treats silence as "keep this"), and put any text you want rendered in quotes rather than describing it loosely.

Using this today

If you are on ChatGPT or the OpenAI API, GPT Image 2.5 is live now - pick Flare for volume work and Sunburst when an image needs to survive several edit rounds. In Karya, GPT Image 2 is available today in the Image Studio for the same kind of text-accurate renders and edits, and Nano Banana Pro and Seedream 5.0 cover the rest of the text-in-image and region-edit workflows this same technique applies to.