OpenAI says its new image model keeps important details during repeated edits and cuts waiting time. New sketch and comment tools also give people clearer ways to guide the result.
New to this? Read it in simple words
- OpenAI released ChatGPT Images 2.5, a new model for making and editing images. It is becoming available in ChatGPT, ChatGPT Work, and Codex.
- OpenAI says waiting time is up to 50 percent lower than with Images 2.0. New sketch, comment, and template tools may help people guide the result.
- Older image tools could slowly change a face or background after several edits. OpenAI says the new model keeps earlier changes more often.
- The 50 percent figure is OpenAI’s own measure, not a promise for every request. Text, hands, logos, and small details can still be wrong.
- Template
- A ready-made design that you can use as a starting point.
Editing is the important change
OpenAI says Images 2.5 creates sharper details and follows editing instructions more closely. It should change only the part a user asks to change. The rest of the subject, layout, and style should remain stable.
This matters during a long creative session. Older image tools could slowly change a face, product, or background after several edits. OpenAI says the new model keeps earlier changes more often and loses less quality over time.
The company also reports up to 50% lower waiting time than Images 2.0. This is an OpenAI measure, not a promise for every request. A complex image can still take longer than a simple one.
OpenAI reports up to 50% lower waiting time. Stable details and clear human control remain the more useful test.
People get more ways to show an idea
A new Sketch tool lets a person draw a rough shape inside ChatGPT. The model can use that drawing as a guide. This can be easier than writing a long description of where every object should go.
Users can also place comments on one part of an image. Templates offer a starting point for work such as posters and product pictures. Prompt sharing lets another person reuse the main idea with different images or details.
These tools move the product closer to a simple design workspace. They do not replace careful review. Text inside an image, real-world facts, hands, logos, and small details can still be wrong or misleading.
Origin signals still need support
Images 2.5 is rolling out in ChatGPT, ChatGPT Work, and Codex. Developers can use two API models. Flare focuses on speed, while Sunburst offers more control for detailed work and takes longer.
OpenAI says generated images include C2PA information and an invisible watermark. These signals can help identify the tool that made an image. They are useful only when websites and viewers keep and show that information.
The practical test is simple. Can a user reach the wanted result with fewer full restarts? Faster output is helpful, but stable edits and clear human control can save more time than speed alone.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3
We used OpenAI for features, access, safety tools, and its speed claim. We checked the early product experience with Axios. We label OpenAI measurements as company reports and do not treat a short test as proof for every image task.