Sketchbook of pencil drawings by Bessie Beebe, 1884 to 1886, Cooper Hewitt, Smithsonian Design Museum. Public domain, via Wikimedia Commons.
Sketchbook of pencil drawings by Bessie Beebe, 1884 to 1886, Cooper Hewitt, Smithsonian Design Museum. Public domain, via Wikimedia Commons.
Sketch is the most useful thing in the new ChatGPT image model, as long as you stop treating the drawing like a prompt. The first public test shows exactly where it breaks.
Every AI image maker has been here. You can see the picture in your head. The subject on the left, the light coming in from the right, a clean empty strip at the top for a title. You write four sentences trying to describe it, hit generate, and get something that is gorgeous and completely wrong.
On September 8, 2026, OpenAI released ChatGPT Images 2.5 with a feature aimed right at that frustration. It is called Sketch, and it lets you draw directly inside ChatGPT and hand the drawing to the model as a visual reference. Stop describing the layout. Just draw it.
It is a lovely promise. The first public test of it came out the same day, and the numbers tell a more useful story than the launch did: a drawing on its own scored 64.6 out of 100. The same drawing with words attached scored 97.8.
The official pitch, as reported by Unite.AI and DataCamp from OpenAI’s announcement:
Developers get two API models. Flare is the default, pitched for speed and volume. Sunburst is the premium option, pitched for tighter control across edits. DataCamp reported leaderboard scores of 1520 for Sunburst and 1491 for Flare in image editing, and 1421 and 1399 in text to image, each with a margin of error of 9 to 13 points.
One more number from the launch puts the stakes in perspective. OpenAI says people already create more than 3 billion images every week across ChatGPT Images and its image models in the API. When a feature changes how people prompt, it changes it at that scale.
Curtis Pyke at Kingy AI published a hands on test on launch day, and it is worth reading because he did something launch coverage rarely does. He kept the failures.
The setup was simple. One ChatGPT session at zero cost, and eight deliberately rough diagrams made locally rather than drawn by hand: a movie poster, a living room, a beverage ad, a sneaker concept, a creator thumbnail, a creature character, a cabin book cover and a website hero banner. Each subject ran under three conditions, and he scored only the first output every time, no rerolls. The scoring gave 40 points to layout match, 25 to whether the model recognized the idea, 20 to detail compliance and 15 to whether you would actually share the result.
Here is what came back.
That is 23 images out of 24 planned runs. The missing one is the “hard limit” in his headline: the sneaker concept with only a sketch returned no output at all after 90 seconds. He left the failure in the results instead of quietly rerunning it, which is exactly what makes the rest of the numbers believable.
The averages hide how badly the worst cases went. With only a sketch to go on, the creature character scored 19 out of 100. Instead of a creature, the model produced an unrelated composition of a face and a cube. The living room scored 34, because the model turned the interior into an exterior.
Think about why. A rough box with a smaller box inside it could be a sofa in a room or a house on a lot. A blob with two dots could be a monster or a portrait. Your drawing is unambiguous to you because you know what you meant. The model does not. It sees shapes and has to guess the noun.
Add one sentence that names the noun and the guessing stops. That is the jump from 64.6 to 97.8.
Pyke’s advice is short and it matches the data. Draw the order of attention. Make the main object larger than everything else. Mark negative space, because empty areas tell the model where a headline or logo belongs. Use arrows and boxes. Then, in his words, “add the brief that tells the model what the drawing should become.”
In practice, for the kind of character art this site is built on, that means:
The honest takeaway is not that Sketch is overhyped. It is that Sketch is a layout tool pretending to be a prompt. Used as a layout tool, paired with a real description, it produced the best average of any condition tested. Used as a replacement for words, it gave up roughly a third of the score and, once, produced nothing at all.
So draw the picture in your head. Then tell it what it is looking at.
OpenAI, Introducing ChatGPT Images 2.5, September 8, 2026; Unite.AI, September 8, 2026; DataCamp, ChatGPT Images 2.5: Features, Sketch, and API Models, September 9, 2026; Curtis Pyke, Kingy AI, ChatGPT Images 2.5 Sketch Test, September 8, 2026. Photo: Bessie Beebe, sketchbook of pencil drawings, 1884 to 1886, Cooper Hewitt, Smithsonian Design Museum, public domain, via Wikimedia Commons.
.jpg){kind=link}