GPT Images 2.5 promises edits that leave the rest of your image alone
The New Stack Meredith Shubel ● Covered by 5 sources
OpenAI’s GPT Images 2.5 is meant to edit one part of a picture without wrecking the rest. The tricky bit: Flare and Sunburst have the same token rates, but not the same speed or control.
Based on reporting by The New Stack, Meredith Shubel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI’s GPT Images 2.5 is aimed at a very specific pain point: making a change to part of an image without forcing the rest of it into a minor identity crisis. That pitch lands hard for anyone who has watched an edit turn a product shot, a background, or a bit of copy into something slightly off everywhere else.
The new release comes with two models, Flare and Sunburst, and OpenAI is drawing a clear line between them. Flare is the default pick, built for everyday work and lower latency. Sunburst is the careful one, meant for premium visual jobs where tighter control matters more than speed.
OpenAI says Flare produces higher-quality images than GPT-Image-2 while cutting latency by 50%. It also says Flare can handle social content, quick prototypes, visual search, and image generation. Sunburst, by contrast, is described as better for production-ready campaigns and product imagery, where precision across edits is the selling point. But the company does not spell out exactly how much longer those extra-precise runs take.
Pricing is also murky in a way developers probably won’t love. OpenAI lists the same token rates for both models: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens. But it gives no clear way to estimate token consumption for GPT Images 2.5, and it explicitly says the GPT Image 2 calculator does not estimate token usage for the new models.
So the basic promise is real enough: more faithful edits, better instruction-following, and less collateral damage when changing an image. But if teams want to know whether Sunburst’s extra control is worth the time cost, or what either model will actually cost on a real job, they’ll have to test it themselves.
My take — AI-written commentary, not fact-checked reporting
This is classic OpenAI: the shiny part is easy to understand, the bill is not. A model that edits one object cleanly is useful; a pricing page that can’t tell you what that clean edit costs is less impressive. The industry keeps selling precision like magic and charging for uncertainty like it’s a feature.
Read more about this at: The New Stack