What shipped
OpenAI released ChatGPT Images 2.5 on 8 September, describing it as its new state-of-the-art image model. It is rolling out to ChatGPT, ChatGPT Work and Codex users across all tiers, on desktop, mobile and web.
The company says the model produces more natural lighting and richer textures, better preserves subjects from a user’s reference photos, follows editing instructions more reliably across several turns of a conversation, and renders images containing real-world information more accurately. It also says generation latency is down by up to 50% compared with Images 2.0.
Those are OpenAI’s own descriptions of its own model, published without an independent evaluation alongside them. Treat them as claims until someone measures them.
Sketch, and the shift it signals
The genuinely new interface piece is Sketch, invoked by typing “@Sketch” in a chat, which lets a user draw directly in ChatGPT and use the drawing as a reference for the final image.
Alongside it come templates for common formats such as flyers and product photos, the ability to leave comments directly on a generated image to steer a targeted edit, and prompt sharing.
Read together, these are not image-generation features. They are document-editing features. The pattern that has held since the first diffusion models — type a prompt, get a picture, type a better prompt — assumes the output is disposable. Templates, region comments and a sketch canvas assume the opposite: that the user has something specific in mind, will iterate on one artefact, and needs to point at a part of it.
The API side
Two companion models went to developers. GPT-Image-2.5 Flare is positioned as the default, which OpenAI says delivers higher-quality images than GPT-Image-2 at 50% lower latency. GPT-Image-2.5 Sunburst is aimed at premium workflows that need tighter control across successive edits.
Both are priced at $8.00 per million tokens for image input and $30.00 per million tokens for image output. Splitting a single image model into a fast default and a controlled premium tier mirrors what OpenAI already does with its text models, and it tells you where the company thinks the money is: not in one-off generations, but in workflows where an image goes through five edits and each one has to preserve the last.
What to watch next
Whether the multi-turn editing claim holds. Preserving a subject across successive edits has been the weak point of every image model to date — backgrounds drift, faces change, text degrades — and it is precisely what the templates and comment features assume works. The first independent tests will be worth more than the launch post.
