OpenAI drops ChatGPT Images 2.5 with 50% faster speeds and Sketch

Trying to describe an exact visual composition using only text has always been one of AI's biggest annoyances. You type a paragraph describing a living room layout, and the model puts the coffee table in midair. Or worse, you ask to tweak a character's shoes, and it completely redraws their face.

OpenAI just launched ChatGPT Images 2.5 on September 8, addressing those friction points directly.

According to OpenAI, the new model cuts generation latency by up to 50% compared to Images 2.0, while delivering more natural lighting, richer textures, and better subject consistency from reference photos. The update is now live across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web—touching a base that already generates over 3 billion images weekly across ChatGPT and the API.

Sketching directly into the prompt

The biggest workflow shift is that OpenAI is finally moving beyond pure text prompting.

Users can now type @Sketch in ChatGPT to bring up a built-in drawing canvas. You don't need any drawing skills. A quick mouse doodle outlining a room layout or an outfit concept is enough for the model to use as a structural guide. Early hands-on testing confirmed that a rough doodle of a cat drawn with a mouse was easily converted into a realistic photo.

Editing also gets much more surgical. Instead of re-rolling the whole image, you can now drop comment pins directly onto specific areas of an image to request targeted changes, like changing an eye color from blue to green while keeping everything else intact.

OpenAI also introduced ready-made templates for posters, merchandise, and flyers, along with the ability to share images with reusable prompts attached. For developers, the API splits into two distinct flavors: GPT-Image-2.5 Flare, which matches the 50% latency drop for high-volume apps, and GPT-Image-2.5 Sunburst, an option that trades longer generation times for maximum precision.

On the LMSYS Arena leaderboards, Sunburst and Flare quickly claimed the top two spots in both generation and single-image editing, pushing previous frontrunner GPT-Image-2 into third. It is worth noting that Arena's sample size for the newcomers remains preliminary, with around 3,000 votes each compared to over 78,000 for GPT-Image-2.

Why it matters

We think this release signals an important pivot: the frontier of image AI is no longer about raw aesthetic flashiness, but about fine-grained workflow control.

For creators and designers, this update is immediately worth testing. The biggest barrier to using AI imagery in client work has always been the lack of predictability. Having to throw away a great composition just because you needed to swap out one background prop is infuriating. Combining quick sketches with pinned comments attacks that problem directly, drastically cutting down on discarded prompts.

For the broader industry, OpenAI is effectively tightening its grip. Holding the top three spots on the Arena boards—even with preliminary votes—forces competitors like Midjourney and Grok to rethink their user interfaces. Simply offering a text prompt bar is no longer enough.

The dual API strategy is also smart product packaging. Developers running social apps or visual search can use Flare for speed and cost efficiency, while marketing and ad production teams can swallow Sunburst's latency in exchange for tight editing control. The real test now is whether the model's multi-turn editing stability holds up under grueling enterprise production demands.