ChatGPT Images 2.5 cuts generation time by half, adds Sketch

OpenAI shipped ChatGPT Images 2.5 on September 8, and the headline number is a 50% cut in generation latency compared to the previous version. Speed alone wouldn’t be much of a story, but the quality jump rides alongside it: textures and lighting read as noticeably more realistic in early samples, and faces or objects pulled from reference images carry over with far less of the drift that made older generations feel like an approximation rather than a match.

The new Sketch feature, called with @Sketch, lets someone draw straight inside the chat window: a rough room layout, a clothing silhouette, whatever loose shape captures the idea, and the model builds finished art on top of that scaffold instead of starting from a text description alone. Editing got more precise too. The model can now change one specific element, a line of text, a background, a single object, while leaving everything else in the image untouched, and it holds that consistency across a long chain of edits rather than degrading a little with each pass the way earlier versions tended to. Templates and prompt sharing round out the release, giving people a faster starting point and a way to reuse setups that already worked.

Developers get two model options through the API instead of one. GPT-Image-2.5 Flare is the fast, default pick built for high-volume generation, carrying the same latency and quality gains as the consumer version. GPT-Image-2.5 Sunburst trades speed for precision, aimed at detailed creative work where the extra generation time is worth the tradeoff. Splitting the API this way lets a developer choose between throughput and fidelity per request instead of picking one model and living with its tradeoffs everywhere.

The rollout is broad from day one. Every ChatGPT, ChatGPT Work, and Codex user gets access across desktop, mobile, and web, with nothing gated behind a separate waitlist or a higher subscription tier this time around. Anyone building prompt libraries around image generation has a fresh set of capabilities to account for here, particularly the reference-fidelity improvements that change what a well-written prompt can actually reliably produce.