On 8 September OpenAI began rolling out Images 2.5 to all ChatGPT, ChatGPT Work and Codex users. Per the company generation latency is up to 50% lower than Images 2.0, and the most interesting addition is the simplest one: you draw roughly in the chat and the model starts from the drawing.
- Sketch: you draw right in ChatGPT, call it with @Sketch, and the drawing becomes a visual guide for the final image.
- Templates for posters, merch and other formats, comments placed on the image itself, and sharing the prompt along with the picture.
- The API gets GPT-Image-2.5 Flare as the default choice for most applications and GPT-Image-2.5 Sunburst for more precise work with longer generation.
Every good briefing I have sat in ended with a doodle on a napkin. Nobody remembers exactly what was said; everybody remembers the drawing.
OpenAI has put the napkin into the chat. The new feature is called Sketch: you roughly draw the room, the silhouette of a garment or just a funny doodle, and ChatGPT turns it into a finished image.
Why the sketch matters more than the speed
You will feel the speed right away and forget it within a week. The sketch changes the way you talk to the machine.
Until now everything went through words. You describe a composition you see in your head, the model reads it its own way, you correct with more words. Anyone who has written a brief knows how much gets lost on the way. A line on paper says where the horizon is, where the person stands and where the light comes from, without arguing over what "left" means.
The second thing worth noting is a quiet one: a comment placed directly on the image. You point at the spot and say what you want there. That is how you work with every designer in the world; ChatGPT just didn't have it until now.
The quality and speed claims are OpenAI's. The real test will come from everyday client work, not from the launch demo.
If you make visuals for clients, try the sketch with something you have already drawn by hand. That is where you will see fastest whether the model listens to the line or only uses it as an excuse.