OpenAI shipped GPT Image 2.5 as two models rather than one, and the split matches how image work actually goes.
Most of the day you want an image now. You are trying a layout, filling a deck, testing whether an idea is worth pursuing at all. Quality matters, but iteration speed matters more, because the tenth attempt is usually the one you keep. That is Flare.
Then there are the few images that carry real weight. The packaging where the type has to be readable. The cutout that has to drop cleanly onto a new background. The edit where one object changes and everything else stays exactly where the photographer left it. That is Sunburst.
Both are live in Kolbo.AI now, in creation and in editing, at 4 credits a generation.
The Same Object, Across Every World
Reference consistency is the thing that decides whether an image model is useful for commercial work or just fun to play with. If you cannot put the same product in twenty different scenes and have it stay the same product, you cannot use it on a campaign.
One cap. A studio shot, a portrait in a doorway, a salt flat at midday, a concrete stairwell. The stitching, the proportions of the peak, the cream stripe and the shape of the numbered patch survive every change of light and location.
You can hand either model up to 16 reference images to hold that line, which is more headroom than we had before and enough to describe a product from several angles at once.
Text That Says What You Typed
Image models have historically been bad at lettering in a way that is hard to design around. Close-but-wrong text is worse than no text, because you notice it a beat after you have already shown the work to someone.
This poster came out of the model with the headline intact, the halftone dot structure consistent across both ink colours, and the number sitting where it was asked to sit. Sunburst is the one to reach for when the lettering is load bearing.
Masked Editing That Leaves the Rest Alone
The most useful thing in the release is also the least dramatic to describe: you can change one region of an image and genuinely keep the rest.
Same photograph, twice. The number on the patch goes from 09 to 24. The steel ruler has not moved, the coffee has the same surface, the cracks in the yellow wall are the same cracks, and the shadow falling across the top left corner lands in the same place at the same angle.
That is the difference between an edit and a re-render. A re-render gives you a new image that resembles the old one. An edit gives you your image back with one thing changed, which is what a revision round actually needs.
Wide Enough for Real Layouts
Fifteen aspect ratios ship with both models, including 21:9, 9:21 and true 3:1 panoramas like the one above. Web heroes, story frames and billboard proportions come straight out of the model instead of being cropped out of a square afterwards, which is where composition usually goes to die.
Native transparent backgrounds are supported too, so a cutout arrives as a cutout rather than as a subject on white that somebody has to mask by hand later.
Every Frame Above Was Made With It
The launch film we cut for GPT Image 2.5 uses no stock, no gallery pulls and no images from another model. One cap, carried across seven worlds, and a montage that runs from editorial fashion through claymation, risograph and 8 bit. Everything you have seen on this page came out of the model being announced on it.
That is the honest test for an image model, and it is the one worth applying to any launch: if the announcement is not made with the thing, ask why.
Which One Should You Use
Reach for Flare when you are iterating. Concepts, layout tests, social variations, anything where you are going to generate a dozen and keep two.
Reach for Sunburst when the image is going somewhere. Packaging, campaign creative, anything with typography in it, and every masked edit where the rest of the frame has to survive.
Both sit in the same model picker, both cost 4 credits a generation, and both offer five quality tiers so you can spend more only on the images that earn it.
What is the difference between Flare and Sunburst?
Flare is tuned for fast everyday image work. Sunburst is tuned for precision: legible typography, exact masked edits and campaign creative you need to get right rather than get quickly.
How many reference images can I use?
Up to 16 on either model.
Can it edit an existing image?
Yes. Both models work in creation and in editing, including masked edits that change one region and leave the rest of the frame untouched.
Does it do transparent backgrounds?
Yes, natively, so cutouts arrive already cut out.
What does it cost?
4 credits a generation, with five quality tiers and resolution options up to 4K.
GPT Image 2.5 is live now. Open Text to Image in Kolbo.AI, pick Flare or Sunburst, and start with the one that matches the job.