ChatGPT Images 2.5: faster generation, precise edits and Sketch

OpenAI released ChatGPT Images 2.5 on September 8, 2026. The headline number is speed: OpenAI says generation latency is up to 50% lower than Images 2.0. The more useful change, though, may be what happens after the first image appears. The new model is designed to preserve a reference subject, change a specific part of an image, and hold earlier details steady through a longer editing conversation.

The release also adds Sketch, image comments, templates and prompt sharing inside ChatGPT. For developers, the same model family arrives as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. This guide separates OpenAI's published claims from the interface details reported in early hands-on coverage.

Want to test the API models now? Start with GPT Image 2.5 Flare for faster iteration, or choose GPT Image 2.5 Sunburst when a detailed final image needs tighter control.

What OpenAI says changed in ChatGPT Images 2.5

The official launch post names four areas of improvement: image fidelity, precise editing, consistency across multiple edits, and better handling of complex visual instructions.

Reference photos are a major part of that story. OpenAI says Images 2.5 is better at preserving recognizable subjects while changing the setting, composition or visual style. It also claims more natural lighting and richer textures. That should matter for portraits and product work, where a polished result is still unusable if a face, package shape or distinctive feature drifts away from the source.

Editing is meant to be narrower. Instead of regenerating the whole composition when you ask for a blue jacket or a different background, Images 2.5 should be more likely to change the named element and leave the rest alone. OpenAI makes the same claim for multi-turn work: earlier edits are more likely to survive later instructions without a steady drop in image quality.

These are vendor claims, not independent guarantees. A selection can still spill beyond the area you marked. OpenAI's own Images in ChatGPT guide warns that highlights are not always precise and edits may extend outside the selected region. Keep the source image and an approved version nearby when an identity, logo or label must remain exact.

Comment on the part you want to change

ChatGPT now lets you place comments directly on an image. A comment gives the model both a location and an instruction, which is less ambiguous than writing “change the object on the left” in a separate prompt. The official editor also supports selecting an area before describing the edit.

In a first-day TechRadar hands-on report, the Edit toolbar included Markup, Comment, Remove BG, Erase and Resize. The reviewer also reported that each edit remained available as a separate version, making it possible to return to an earlier result. Those toolbar labels and version-history details come from that test; OpenAI's help page currently documents selection editing plus Resize and Remove on mobile, but does not present the same five-item list for every platform.

A practical editing request should still say what must not change:

Change only the selected jacket from black to cobalt blue.
Keep the face, pose, hands, lighting, crop and background unchanged.

One focused instruction per turn makes drift easier to spot. Compare the entire frame after each edit, not just the selected object.

Sketch turns a rough drawing into a visual reference

Some layouts are easier to draw than describe. Type @Sketch in ChatGPT, draw a rough composition, then add the subject, colors and style you want. OpenAI's Sketch announcement on X shows the feature as a direct way to give the model spatial guidance.

OpenAI's help documentation currently describes the drawing flow specifically for the ChatGPT mobile app: type @, select Sketch, draw, confirm, add instructions and send. The model uses the sketch as a reference for the final image. A rectangle, arrows and a few labels can be enough to show relative placement; the sketch does not need to look finished.

Templates offer another starting point. ChatGPT includes formats such as posters and product photos, then asks for the copy, design elements or style needed to customize the result. Prompt sharing lets someone else reuse the idea with their own photos and details.

Flare and Sunburst bring Images 2.5 to the API

OpenAI released two API models alongside the ChatGPT update. Their roles are deliberately different.

Sunburst trades speed for additional precision, so it is not automatically the better choice. Start with Flare when iteration speed matters. Move to Sunburst when a final asset needs more careful control and the longer generation time is acceptable.

You can try both through Yix: GPT Image 2.5 Flare is the faster default, while GPT Image 2.5 Sunburst is aimed at detailed final work. The pages show the currently supported controls before you generate.

Availability and the “three billion images” figure

ChatGPT Images 2.5 is rolling out across desktop, mobile and web to users of ChatGPT, ChatGPT Work and Codex on all tiers. The exact interface can differ by platform and rollout stage. For example, OpenAI documents Sketch on mobile, while its launch post describes the broader Images 2.5 release across all three surfaces.

OpenAI also says people create more than 3 billion images every week across ChatGPT Images and GPT-Image models in the API. That is a weekly usage figure, not a lifetime total of three billion.

OpenAI co-founder Greg Brockman quoted the launch announcement. Sam Altman's release post joked that Images 2.5 probably cannot solve “super difficult math problems,” before calling the model very good. The joke landed on a busy day for OpenAI, but it also keeps the product boundary clear: this is an image generation and editing model.

What to test first

The fastest useful test is not a spectacular text-to-image prompt. Use a reference image you know well, make one small edit, then make a second edit somewhere else. Check whether the first change survives and whether untouched details move. That exposes the two promises at the center of Images 2.5: local precision and multi-turn consistency.

For production work, keep an approved master, edit one concept at a time and inspect faces, hands, typography, product labels and edges after every pass. Faster generation helps, but fewer silent changes are what make an image editing workflow easier to trust.