What is new in ChatGPT Images 2.5?
OpenAI introduced ChatGPT Images 2.5 on 8 September 2026 as its new state-of-the-art image model. The headline changes are not limited to prettier outputs. The update focuses on the parts of image generation that become important once you start refining a real asset: preserving a person or product from a reference image, changing one element without unnecessarily rebuilding the rest, and keeping earlier edits intact over a longer conversation.
OpenAI says Images 2.5 produces more natural lighting and richer textures, is better at preserving subjects from reference photos, follows complex visual instructions more reliably, and can handle more demanding layouts including transparent backgrounds. It also reduces image-generation latency by up to 50% compared with Images 2.0. That figure is a maximum reported reduction, not a promise that every prompt will finish twice as fast.
| Area | What Images 2.5 changes |
|---|---|
| Reference fidelity | Subjects from reference photos are more likely to stay recognizable across new settings, styles, and compositions. |
| Targeted editing | The model is better at changing the requested element while leaving unrelated details alone. |
| Repeated edits | Earlier changes are more likely to remain consistent through multiple editing turns without progressive quality loss. |
| Prompt understanding | Complex visual directions, composition requirements, and style constraints are handled more reliably. |
| Speed | OpenAI reports up to 50% lower image-generation latency than Images 2.0. |
| Creation tools | Sketch, Templates, image comments, and prompt sharing add new ways to start and refine images inside ChatGPT. |
The biggest practical change is more precise image editing
Older image-editing workflows could force a tradeoff: fix the one thing you dislike and risk changing several things you wanted to keep. Images 2.5 is designed to reduce that problem. OpenAI says the model is better at editing only what you ask it to edit while preserving the subject, composition, background, and surrounding treatment.
That matters most when an image is already close to finished. Imagine a product photo where the object, lighting, framing, and brand treatment are correct, but the label color is wrong. A useful editor should change the label color without casually changing the product shape, camera angle, or typography. The same principle applies to portraits, posters, thumbnails, room mockups, social graphics, and illustrations.
Images 2.5 also improves reference-led transformations. OpenAI says recognizable features from reference photos are more likely to carry through when the subject is moved into a different setting, composition, or visual style. This does not mean every identity, logo, piece of text, or small detail will be preserved perfectly. It means preservation is an explicit area of improvement in the new model.
Useful prompt pattern: name both the change and the parts that must stay fixed. For example: “Replace only the background with a clean studio wall. Keep the person’s face, pose, clothing, lighting direction, crop, and camera perspective unchanged.” Images 2.5 is better at this kind of focused instruction, but a precise request still gives the model a clearer target.
Official examples and description: OpenAI’s precision-editing section.
Multi-turn editing should drift less as you keep refining
Image creation is often iterative. You generate a base image, fix a hand, change a headline, adjust the background, move an object, and then make one final lighting change. The difficulty is that every new instruction gives an image model another opportunity to alter details you had already approved.
OpenAI describes multi-turn editing consistency as a core Images 2.5 improvement. Earlier changes are more likely to remain in place while later edits build on the same asset, and the image is less likely to degrade simply because you have made several consecutive revisions.
For creators, this can change how you work. Instead of trying to write one enormous prompt that anticipates every detail, you can start with the strongest overall composition and then make narrower passes. The better the model becomes at remembering what should remain stable, the more practical conversational editing becomes.
Sketch, Templates, comments and prompt sharing add new ways to work
Images 2.5 arrives with several ChatGPT product features that change how you communicate visual intent. They are useful for different stages of the workflow rather than being four versions of the same tool.
- Sketch
- On mobile, type
@in the message box and choose Sketch. You can draw a rough idea and then describe the finished image you want. The sketch acts as a visual reference, which is useful when spatial relationships are easier to draw than explain. - Templates
- Open Images, choose Templates, select a starting format, and customize the details. OpenAI gives examples such as posters and merch. Templates are not yet available in ChatGPT Work during the current rollout.
- Image comments
- On mobile, opening a generated image full-screen can expose editing and commenting controls. Comments let you point to a specific place in the image and describe the change you want more locally.
- Prompt sharing
- You can share the prompt behind a generated image so another person can try the same concept with their own images and details rather than only viewing the final output.
These features reduce different kinds of friction. Sketch helps when words are not enough. Templates reduce blank-page setup. Comments make local edits easier to communicate. Prompt sharing turns a finished idea into something reusable.
Current product details: OpenAI’s ChatGPT release notes for 8 September 2026.
Who can access ChatGPT Images 2.5?
OpenAI says Images 2.5 is rolling out to ChatGPT, ChatGPT Work, and Codex users across all tiers, and that the rollout covers desktop, mobile, and web. You should not need a specific paid ChatGPT tier merely to be part of the Images 2.5 rollout.
That does not mean every account will display every new control at exactly the same moment. “Rolling out” means availability can arrive progressively, and some features have surface-specific limitations. OpenAI’s release notes specifically say that Templates are not yet available in Work mode. They also describe Sketch and the new full-screen edit/comment experience as mobile features.
OpenAI also says existing image-generation limits are unchanged. In other words, the 2.5 model launch does not itself announce a new universal quota. The amount of image generation available to you can still depend on your plan, product, account state, and any limits shown in the interface.
| Question | Current answer |
|---|---|
| Which ChatGPT tiers? | OpenAI says the rollout covers all ChatGPT tiers. |
| Which products? | ChatGPT, ChatGPT Work, and Codex. |
| Which platforms? | Desktop, mobile, and web. |
| Are Templates in Work? | Not yet, according to the current release notes. |
| Did image limits increase? | No new universal limit was announced; OpenAI says existing image-generation limits are unchanged. |
Sources: OpenAI pricing and availability section and ChatGPT release notes.
Developers get two new GPT-Image-2.5 API models
The ChatGPT product name and the API model names are not identical. For developers, OpenAI introduced GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
Flare is positioned as the default choice for most applications. OpenAI says it carries the new quality, editing, and speed improvements into the API and delivers higher-quality images than GPT-Image-2 at 50% lower latency. OpenAI suggests it for creator content, social content, product experiences, visual search, rapid prototyping, and higher-volume generation.
Sunburst is aimed at premium visual workflows where tighter control is more important than speed. OpenAI describes it as suitable for production-ready campaign creative, polished product imagery, and other detailed editing workflows, with longer generation times.
The important distinction is that ChatGPT access does not automatically describe how a developer should build an API workflow. If you are integrating image generation into an application, use the current API documentation and pricing for the API model you actually select.
Official model positioning: OpenAI’s Images 2.5 API section.
A better way to edit images with Images 2.5
The model is more capable, but the editing process still improves when each instruction has one clear job. A controlled workflow makes it easier to see which change helped and which change introduced a new problem.
- Start with the strongest base image. Do not spend ten edits rescuing a composition you already dislike. Get the framing, subject, and overall visual direction reasonably close first.
- Make one meaningful edit at a time. Ask for the background change, then the text change, then the wardrobe change instead of combining unrelated corrections into one ambiguous instruction.
- State what must remain unchanged. When preservation matters, name the face, pose, product shape, crop, typography, lighting, or other details you want locked.
- Use location language. Say “the small label in the upper-right corner” or use the selection/comment tools when the change belongs to one specific area.
- Refer to the approved state. In longer conversations, say which previous features should stay: “Keep the background and clothing from the current version; change only the headline.”
- Stop and regenerate when the foundation drifts. Better multi-turn consistency reduces drift; it does not make every edit reversible or perfect. If several important details have moved away from the target, returning to a cleaner base can be faster than stacking corrections.
ChatGPT’s image editor also supports direct instructions and area selection. OpenAI notes that selections are not always exact and an edit can extend beyond the highlighted region, so textual constraints still matter even when you use the selection tool.
Editor mechanics: OpenAI Help Center, Images in ChatGPT.
What Images 2.5 improves without guaranteeing
“More precise” does not mean pixel-locked. “Better reference fidelity” does not mean every facial feature, logo, font, hand, or tiny object will survive every transformation unchanged. “Up to 50% lower latency” does not mean every request will be 50% faster. And better multi-turn consistency does not remove the possibility of drift after repeated edits.
The most useful way to understand Images 2.5 is as an improvement in control and iteration. It should require fewer compromises when you already know what you want to preserve, and it gives you more ways to communicate that intention through text, selections, comments, sketches, templates, and reference images.
OpenAI says the release continues to use prompt and image safety checks, C2PA metadata, and invisible watermarking to help identify content made with its image tools. Those safeguards are part of the model’s deployment rather than optional editing features.
If you are deciding whether the update matters for your own workflow, test it on the task that used to cause the most frustration: a portrait that must stay recognizable, a product image with one small change, a poster that needs repeated revisions, or a multi-step edit where previous models tended to drift. That is where the difference between a better generator and a better creative tool becomes easiest to see.
Safety and model claims: OpenAI’s ChatGPT Images 2.5 announcement.