ChatGPT Images 2.5 is twice as fast and holds its edits

The biggest change is in how you make corrections. You can now leave a mark directly on the image, like a comment, and only that spot gets changed. The new sketch feature is a tool for roughly drawing the layout or composition you want by hand and feeding it in as a reference. It hands over position and size, which are hard to describe in words, as a drawing.

The developer API ships with the model split in two. Flare is the default model: half the latency of GPT-Image-2 with higher quality, tuned for creator content and high-volume generation. Sunburst is the upper model, holding detail more firmly across repeated edits and producing cleaner transparent backgrounds, for finishing work such as ad comps and product images. According to OpenAI's official video, both models support transparent backgrounds and Flare is more than 50% faster.

Seen through actual use, the two models split like this. A team pulling hundreds of images a day to fill thumbnails or product listings belongs with Flare, while a designer revising a single ad comp ten times to finish it belongs with Sunburst. By deliberately splitting the name in two, OpenAI has dropped the promise of doing both jobs at the same price and is selling speed and precision separately.

There are things that came bundled with it. Templates for frequently used formats have been added, and prompts can now be shared along with the image. Generated images carry C2PA provenance information and an invisible watermark.

The competitive landscape in image generation moves with it. In a market where Google's Nano Banana line and the open models hosted on fal have competed on speed and price, OpenAI has this time halved its speed while putting editing stability out front. There are now plenty of models whose first image is pretty; the stage now is choosing the model that stays intact after you fix it.

The change that reaches Korean users right away is editing. Until now it was common to try to fix one spot, have the whole image change, and start over from scratch; comment editing and consistency that survives repeated revisions cut that loop. The sketch feature is especially useful for people who found it hard to describe a composition in Korean.

There is a direct effect on METAL's own pages. Since this week, the hero images for our business articles have been made with the GPT-Image-2 family, and Flare and Sunburst become the next options for that slot. Flare fits automated publishing where speed matters, and Sunburst fits hand work that goes through several rounds of editing.

Price has to be looked at separately. OpenAI did not state the fees in its announcement and pointed to the developer page instead, and outlets such as DataStudios reported that the API price is around twice that of the previous generation. Speed halved, but the price per image went up, so teams using it in volume should first work out how much Flare saves them.

In the end, what this round is about is not new images but the cost of fixing them. There are already several models that get it right the first time, but models that do not break when you fix them were rare. OpenAI has started selling that spot as two models, and the image generation race is now decided not at the first image but at the tenth revision.