OpenAI started rolling out ChatGPT Images 2.5 on Tuesday, to ChatGPT, ChatGPT Work and Codex users across desktop, mobile and web.
What changed
Speed is the part you feel: up to 50 percent lower latency than the previous version. On top of that, sharper detail, better lighting and texture, and reference photos that keep their subject instead of turning into someone else after two edits. That last one is exactly where the previous generation kept falling over.
Three new tools come with it. A sketch feature lets you draw directly in ChatGPT and hand the scribble over as visual reference. There are templates for common formats. And you can share prompts. Edits can now be driven through comments as well.
For developers
The API gets two variants:
- GPT-Image-2.5 Flare is the default. Higher quality at half the latency, aimed at social content and high volume.
- GPT-Image-2.5 Sunburst targets precise editing control — campaigns, product imagery, anything that actually ships.
OpenAI doesn’t name prices in the announcement.
Reading it
Splitting into a fast model and a precise one is the real news here. Image generation has long been a single dial between “quick” and “good”, and you paid for both no matter which you needed. Batch out thumbnails and you don’t want Sunburst quality. Retouch a product shot and Flare’s speed does nothing for you.
For me it still comes down to consistency across edits. Generating one good image in a single shot is table stakes now. Recognizing the same face after the fourth “make the background brighter” is the job nobody has nailed. I’ll test that the moment it shows up here.
Sources: OpenAI: Introducing ChatGPT Images 2.5, 9to5Mac, Axios