The speed boost is the least interesting thing about ChatGPT Images 2.5. That will not be the headline you see elsewhere, because “50% faster” is easy to put in a chart and easy to cheer for. But if you are someone who has ever tried and failed to describe a picture in words, the real story of OpenAI’s September 8, 2026 release is a small drawing box called Sketch. Speed makes an existing habit less annoying. Sketch changes what you are allowed to bring to the conversation.
What actually shipped
OpenAI released ChatGPT Images 2.5 and called it the company’s new state-of-the-art image model. The upgrade targets the complaints people had about Images 2.0, chiefly slow generation and subjects that morphed into slightly different people or objects between edits. Generation latency is down by as much as 50% compared with Images 2.0, and the company points to sharper details and more precise editing.
Alongside the ChatGPT experience, two companion models arrived in OpenAI’s API:
- GPT-Image-2.5 Flare — matches the same quality, editing, and speed gains, and is the default for most apps.
- GPT-Image-2.5 Sunburst — adds more precision for detailed editing work.
There are also templates for popular formats, which is a quiet quality-of-life addition for anyone who has ever guessed at the right aspect ratio for a social post.
Why Sketch is the part worth your attention
Sketch is invoked with @Sketch, and it lets you draw directly inside ChatGPT. Your drawing then works as a reference for image generation. That is the whole feature. It sounds minor. It is not.
Think about what prompting an image model has required until now. You had a picture in your head, and the only channel to get it out was language. So you typed things like “a coffee shop interior, warm light from the left, the counter running diagonally toward the back right corner, a person seated in the lower third facing away.” Then you read the result, noticed the counter was in the wrong place, and tried again with more adjectives. Prompting for images has been a game of describing spatial relationships in a medium that is bad at spatial relationships.
A drawing skips all of that. Three rectangles and a stick figure communicate layout instantly, and they do it without requiring you to know a single art term. You do not need to know what “rule of thirds” or “three-quarter view” means. You need to know roughly where you want things. Most people can do that with a mouse and thirty seconds, even if what they produce looks like it was made under duress.
That is a real shift in who gets to direct these tools. The people who got the most out of image models were often the ones with the vocabulary for it, whether from design training, photography, or just a lot of practice. Sketch lowers the entry cost for everybody else.
What the two API models mean for AI agents
If you follow this site, you probably care less about generating a birthday card and more about what your automated helpers can do. The two API models are the part that matters there.
When an AI agent needs to produce or edit an image, some model has to do that work behind the scenes. Flare being the default for most apps is a sensible signal: it is the general-purpose option, fast and good enough for the vast majority of jobs. Sunburst exists for the cases where fine detail is the point, like careful edits to an existing image where small errors are obvious.
Practically, this gives the people building agents a choice they did not have to think much about before. An agent that generates a hundred product thumbnails overnight wants Flare. An agent that touches up one photo a customer is going to look at closely wants Sunburst. That kind of tiering is unglamorous, and it is exactly how these tools become dependable enough to hand real work to.
The latency cut compounds this. Agents rarely do one thing. They do a step, evaluate it, and do another. Cutting the wait on each image step by up to half changes what is comfortable to build. A workflow that generates, checks, and regenerates was tedious at the old speed and becomes reasonable at the new one.
The pattern underneath all of this
OpenAI has said people create more than 3 billion images with its tools, so this is not a niche feature area. And the direction of travel in this release is consistent: give people more ways to say what they want, and make the loop between asking and seeing fast enough that iteration feels natural.
Sketch fits that. So does the templates addition. So does splitting the API into a fast default and a precise specialist. None of it is flashy. All of it reduces the gap between the picture in your head and the picture on your screen.
If you have avoided image generation because you never found the right words, this is a good week to try again. Open ChatGPT, type @Sketch, and draw something ugly. That is now a valid way to ask.
🕒 Published: