OpenAI Launches ChatGPT Images 2.5 With Sketch-to-Image Tool and 50% Faster Generation

OpenAI released ChatGPT Images 2.5 on Tuesday, introducing a new Sketch feature that lets users draw doodles directly in ChatGPT to guide image generation. The update is live now for ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web, according to OpenAI.

Sketch is activated by typing "@Sketch" into the chat box, which opens a drawing window. Users can sketch a rough shape—a cat, a building, an object—and then instruct ChatGPT on how to render it into a finished image. The Verge's Jay Peters tested the feature, successfully converting a mouse-drawn cat doodle into a photorealistic image, then using a new inline commenting tool to change the cat's eye color from blue to green by annotating that specific part of the generated image.

The inline comment feature allows users to mark up specific regions of a generated image with notes on what to change, rather than issuing a fresh text prompt for the entire image.

Alongside Sketch, OpenAI says the underlying Images 2.5 model produces "more natural lighting and richer textures" than its predecessor and better maintains coherence across multi-turn editing sessions. OpenAI also claims image generation latency has dropped by up to 50% compared with Images 2.0, though the company has not published independent benchmark data or a technical report to substantiate that figure.

OpenAI has not disclosed pricing changes, context window specifications, or a training cutoff date for the Images 2.5 model. The rollout appears to be a direct swap for the previous image generation model within ChatGPT rather than a separate opt-in product.

What this means

Sketch closes a gap between text-only prompting and the kind of visual reference control that tools like Midjourney's image-to-image workflows or Adobe's generative fill have offered for longer. For users who struggle to describe a composition in words—an odd camera angle, a specific pose, an unusual object arrangement—drawing it first and refining with text is often faster than iterating through prompt rewrites alone.

The inline commenting system is arguably the more consequential change. Text-based image editing tools have historically struggled with localized edits: asking a model to "make the eyes green" without disturbing the rest of the image often triggers unwanted changes elsewhere in the frame. Letting users point directly at the region they want altered narrows the model's task and should reduce the trial-and-error cycle common in AI image editing.

The claimed 50% latency reduction, if it holds up under independent testing, matters for anyone using ChatGPT image generation in production workflows—marketing teams, game asset prototyping, rapid concept iteration—where wait times between edits compound quickly. OpenAI has not detailed whether this speedup comes from a smaller model, better serving infrastructure, or architecture changes, which limits how much can be concluded about the model itself versus deployment engineering.