ChatGPT Image Editing After DALL·E: What Changed

ChatGPT image editing has become a different product in a few months. On 30 August 2026, OpenAI retired the official DALL·E GPT from ChatGPT. Nine days later it released ChatGPT Images 2.5, which made editing noticeably more precise, and on 4 November Images 2.6 was reported to add animated GIFs.
We first published this post in July, ahead of the retirement, and updated it in November 2026. Below: what actually went away, what the current image system can do, how editing works now, and the limitations OpenAI’s own help documentation still admits to.
What changed on 30 August
What retired was the official DALL·E GPT, the separate assistant you’d open from the GPT list. Two things were not affected: user-created GPTs with image generation enabled kept working, and image generation inside ChatGPT carried on through ChatGPT Images.
Before the cutoff, OpenAI advised users to download any DALL·E GPT images they wanted to keep, but it didn’t spell out what happens to those old conversations afterwards. If you missed the date, open your old DALL·E conversations and see what’s still there before assuming anything is gone.
What replaced it
ChatGPT Images is now the only image system in ChatGPT, and it has moved fast. Images 2.0 arrived in April 2026, built on the gpt-image-2 model, and Images 2.5 followed on 8 September. OpenAI says 2.5 is available on all ChatGPT tiers and brings:
- Faster generation, with latency cut by up to 50% compared with Images 2.0.
- More precise editing. OpenAI says it’s better at changing only what you asked for while keeping the rest of the image the same.
- Steadier multi-turn edits. Earlier changes are more likely to survive as you keep editing, without the image degrading.
- Better reference photos. Subjects from your uploads stay more recognizable in new settings and styles.
- Sketch. Type “@Sketch” to draw a rough layout in ChatGPT and use it as a guide for the final image.
- Templates for common formats like posters and product photos, so you don’t start from a blank prompt.
The improvements from 2.0 are still there too: much better text rendering, aspect ratio control by picker or prompt, transparent backgrounds on request, and an automatic library of everything you create under Images on web and mobile.
Images 2.6, reported on 4 November 2026, adds animated GIF creation from a prompt or an uploaded photo. OpenAI’s full documentation for it wasn’t out when we updated this post, so treat those details as early.
How ChatGPT image editing actually works
YYou can now upload an existing photo, not just images you generated, and change it by describing what you want. There are now three ways to do it:
- Selection-based. Open the image in the editor, use the Select tool to highlight an area, then describe the change in chat. Undo and Redo let you refine the selection before committing.
- Conversational. Skip the selection entirely and describe the edit in the conversation panel. The model works out where to apply it, which is faster but broader.
- Comments on the image. New with Images 2.5: place a comment directly on a spot in the image and describe the change there. It’s a lighter way to point at an area than drawing a selection.
The editor itself is deliberately minimal: Select, Aspect ratio, Undo, Redo, Cancel, Save. On mobile you tap an image to open the editor, choose Edit to describe changes, or Select to highlight an area with a size slider before describing what should happen there.
The interesting part is the interaction model. This is editing as conversation: make one change, look at it, ask for the next. It suits people who know what they want but not which tool would do it, which is a large group that traditional editors have never served well.

The limitations OpenAI admits (and the ones it doesn’t)
OpenAI says Images 2.5 is better at editing only what you ask for, and in practice that narrows the gap. But its help documentation still carries the same caveat: “Highlights are not always precise, and edits may extend beyond the area you selected.” That’s a significant caveat for anyone expecting mask-level control. You are describing intent, not defining boundaries, and the model decides how far the change travels.
Three more constraints worth knowing before you build a workflow on it:
- Deleting an image means deleting the conversation. There’s no per-image delete in the library; you remove the chat that produced it. Awkward if one conversation contains work you want and work you don’t.
- Generation isn’t instant. Images 2.5 is up to 50% faster than 2.0, but complex images can still take a while. You can keep using ChatGPT while it works.
- No exact specifications. You can request an aspect ratio, but not “exactly 2000 × 2000 px, under 1 MB, sRGB.” For anything with hard requirements, like marketplace product images, this remains an editor’s job.
Writing edit instructions that actually work
Conversational editing rewards a different kind of instruction than image generation does. When you generate, you describe a destination. When you edit, you describe a change, and the model has to work out what stays untouched. Four habits make that reliable:
- Name what shouldn’t change. “Replace the background with a plain grey studio wall, keeping the subject, lighting, and shadows exactly as they are” outperforms “put her on a grey background,” because the second leaves the model free to reinterpret the person too.
- One change per message. Batched requests (“remove the cup, brighten it, and make it square”) produce compromises. Sequential edits let you keep what worked and retry only what didn’t, which is the whole advantage of the conversational model.
- Anchor edits to visible landmarks. “The cable on the left edge of the desk” beats “the cable,” especially given that selections spill. Describing position narrows the blast radius even when your highlight doesn’t. With Images 2.5 you can also drop a comment right on the spot.
- Say what the result should look like, not what tool to use. “Make the shadows softer, as if lit through a window” works; “apply a 40% feather with reduced contrast” doesn’t, because there are no sliders behind the curtain.
When an edit goes wrong, resist the urge to correct it with another edit on top. Two or three stacked corrections usually drift further from the original than starting again from the source image, the same principle behind not re-editing AI artifacts indefinitely: at some point regenerating beats repairing.
Where it fits your workflow
The honest summary: ChatGPT image editing is excellent at “make this look different” and poor at “make this exactly right.” Those are different jobs, and most projects need both.
|
Task |
Best tool |
Why |
|
Change a background or scene |
ChatGPT |
Describing beats masking |
|
Remove an unwanted object |
ChatGPT |
Generative fill handles it well |
|
Resize to an exact pixel spec |
Editor |
Deterministic, repeatable |
|
Compress to a file-size target |
Editor |
You need a number, not a guess |
|
Crop, straighten, adjust color |
Editor |
Seconds, with no re-rolls |
|
Batch the same change across 50 files |
Editor |
Conversation doesn’t scale |
|
Add grain or a film treatment |
Editor |
Consistent across a set |
The pattern that works: use ChatGPT for the creative transformation, then finish in an editor for the specifications. Generate or alter the scene, then crop, resize, compress, and export to whatever the destination demands. We covered that division in more depth in AI generators versus photo editors, and the arrival of conversational editing sharpens rather than changes it.
Finish the job with exact control
Crop, resize, compress, and export to spec. Free, in your browser, no re-rolls.
Open the ArtsFlick Photo EditorOne more consideration, given how audiences are responding to AI polish: the easier conversational editing becomes, the more heavily edited images will flood every feed, and the more valuable visibly real photography becomes. Use the tool, but don’t let everything you publish look like everyone else’s output.
