OpenAI's ChatGPT Images 2.5 lets you sketch prompts by hand

OpenAI rolled out ChatGPT Images 2.5 on September 8, 2026, and the headline feature is a drawing pad. Type "@Sketch" into the chat box and ChatGPT opens a canvas where users can doodle a rough shape, then feed that sketch into the model as part of an image prompt.

It is a small interaction change with a real practical use: describing composition in words is hard, but dragging a circle where you want a moon or a box where you want a building is not. The Sketch tool aims to close that gap between intent and output.

Beyond sketching, the update focuses on consistency across a conversation. OpenAI says the model now follows instructions more reliably across multiple back-and-forth turns, so a user refining an image over several messages should see fewer dropped details or reset compositions.

Editing gets more direct too. Instead of describing a change in text, users can now click on a specific part of an image and leave a comment there, pointing the model at exactly what to fix rather than re-explaining the whole scene.

Faster, and better at faces

OpenAI is claiming up to a 50% cut in generation latency compared to Images 2.0, alongside more natural lighting and richer texture rendering in outputs. The company also says the model is better at preserving the identity of subjects pulled from reference photos, a common pain point in earlier versions where faces or objects would drift after edits.

That usage figure gives some sense of scale: OpenAI's image tools, spanning both the ChatGPT consumer product and the GPT-Image models available through the API, have collectively produced more than 3 billion images.

On the developer side, OpenAI introduced two new model IDs, gpt-image-2.5-sunburst and gpt-image-2.5-flare, giving API users access to the same underlying improvements outside the chat interface.

The rollout covers ChatGPT, ChatGPT Work, and Codex, and is available across desktop, mobile, and web. There are no reported contradictions across sources on the launch date, feature set, or availability, and the update reads as an incremental but concrete refinement of OpenAI's existing image pipeline rather than a new standalone product.