OpenAI Releases ChatGPT Images 2.5, Cutting Generation Latency by Up to 50%

OpenAI's technical blog listed four main changes: generation is faster, with latency down as much as 50 percent; images look more natural, with people and objects taken from a reference photo more likely to keep their original features; edits made over several consecutive rounds stay consistent; and users can annotate directly on an image to tell the model what to change. QbitAI also pointed to a sketch function, in which a user types @Sketch in the chat box, draws a rough composition with a mouse, adds style and detail instructions, and the model treats the drawing as a visual reference. New ready-made templates and a percentage-based generation progress indicator were added as well.

The resolution ceiling remains 3840 pixels, but Images 2.5 introduces xhigh and max quality tiers aimed at better detail, material texture, lighting and fidelity to the reference subject. In one demonstration, a child's clothing was replaced while the face barely shifted and the slightly frizzy texture of the original curly hair was largely retained. A second example, in which a dog is dressed in clothes, was less clean: the shadow behind the dog kept the silhouette of its original back, which should have smoothed out once the clothing was added. A demonstration that restored and enlarged an original photograph while reworking its lighting produced a sharper result.

The company is placing more emphasis on controllability than on making pictures simply look prettier. A set of e-commerce demonstrations walked through repeated clothing and background swaps on the same subject; the person, the spatial relationships and the overall composition stayed largely stable, although some frames retained a slightly plastic AI look. QbitAI noted the change in the ratio between figures and objects between Image 2 and Images 2.5.

Annotation-based editing is meant to work the way designers comment on a draft. Instead of writing a long instruction such as changing the sunglasses on the second person at the upper left, a user circles the area and types a short request; in QbitAI's test on a meme-generation screen, marking the sunglasses and writing "change to black-framed glasses" was enough for the model to identify the target. The capability is presented as most useful in images containing many people or objects. Local editing was also strengthened, with a travel-ticket demonstration showing a fixed layout that could be updated city by city while keeping earlier design decisions.

Multi-turn consistency is the third area of focus. Design work rarely ends with one generation: a subject is produced first, then the layout changes, then the text, then fine details. OpenAI says Images 2.5 can carry a longer conversation forward while preserving content and image quality that had already been approved, a scenario tested with a travel-guide poster in which text, photos, routes and layout all occupy the same frame.

Understanding of complex scenes, materials and style instructions was upgraded alongside editing. OpenAI published examples including shattered glass, a Japanese lifestyle magazine cover and a science-fiction scene in the style of Blade Runner. QbitAI's own tests included a set of travel photos generated from a single picture of a monkey, and it observed that some areas still showed a familiar AI smudging effect. Other users produced stop-motion animation from the model and reported sharper detail. Hand-drawn sketch tests followed the general direction but sometimes got proportions wrong, while material and volume rendering improved noticeably, as seen in a shoe whose leather, sole, creases and three-dimensional form looked firmer than before. Some users reported the opposite trade-off, saying detail improved while human limbs failed more often.

OpenAI also released two API models. GPT-Image-2.5 Flare balances quality, editing ability and speed and is the recommended default for most applications, including social content, rapid visual prototypes and large batches of images. GPT-Image-2.5 Sunburst targets finer creative and editing work such as advertising assets, product photography and other high-quality visual content. Both are billed at the same token rates: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. Images 2.5 arrives about four and a half months after GPT Image 2.

Editor's Summary

OpenAI released ChatGPT Images 2.5 with up to 50 percent lower generation latency, in-image annotation, stronger local editing and steadier multi-turn consistency, supported by two API models billed at the same token rates. The update targets design and e-commerce workflows rather than one-off image generation. Early user tests show improved materials and detail but also lingering flaws in proportions, limbs and residual AI artifacts.