On August 30, 2026, OpenAI switches off the DALL-E GPT inside ChatGPT. The instruction that came with it is one line long: download the images you want to keep, then use ChatGPT Images from now on. Six days from today, a tool that made a very large share of the AI pictures on the internet stops answering.
The coverage is doing what coverage does. Save your stuff, here is the export button, here is a list of alternatives. All correct. All beside the point.
The point is that a company put a date on a door, and several million people are about to find out that the only copy of their work was on the other side of it.
What the DALL-E retirement actually removes
Set the argument aside and read the release honestly, because it is not a scandal.
What goes away is the DALL-E GPT, the dedicated image tool inside ChatGPT. What stays is ChatGPT Images, the newer built-in generator running gpt-image-1 and gpt-image-1-mini, which follows instructions better, handles text inside pictures better, and can edit an existing image rather than only make a new one. Custom GPTs that people built with image generation switched on keep working. Users got roughly thirty days of notice.
The developer side already happened and almost nobody noticed. The DALL-E 2 and DALL-E 3 model snapshots were removed from the API on May 12, 2026, with gpt-image-2, gpt-image-1 and gpt-image-1-mini named as the replacements. This month is the consumer half of a decision that was made a year ago.
I should say plainly that I have nothing in that folder. I never used DALL-E. Not once on a paying job. My still image work goes through Midjourney, and inside Higgsfield through GPT Image and Nano Banana Pro, which is where I need text to survive inside a frame and a character to survive across frames.
So I am not writing this as somebody about to lose something. I am writing it because the people who are about to lose something are, almost by definition, the people least likely to have a second copy.
Why is a shutdown date good news?
Here is my contrarian read, and it is not the one the panic invites.
Everybody is filing this under platform risk. Another reminder that the tools are rented. Fair enough. But a retirement announced a month ahead, with a written date and a written instruction to go and download your work, is the most honest thing a model company will do all year.
What should actually frighten a working person is the version with no date on it. The silent revision. The same prompt, the same settings, a different picture next month, no announcement, no changelog, no way to get the old behaviour back. That happens constantly and it never makes the news, because there is nothing to photograph. A shutdown at least has the decency to be visible.
I have argued this from the other direction before, when the Sora API got a sunset date. The dependency is not the problem. The dependency is the deal. The problem is pretending you have not made it.
The thing you lose is not the picture
Be honest about the images for a second. Nobody is going to grieve a 2023 DALL-E render. They are dated in a way that is almost charming now, and any current model will beat them in ten seconds from a worse prompt.
What dies on August 30 is not the picture. It is the record.
The prompt that produced it. The eleven attempts before it that did not work, and why. The order you tried things in. The date. The version of the model that was live that week. That is provenance, and it is the only part of AI image work that cannot be regenerated, because it is a history rather than an output.
Post-production has known this forever. When a job is archived, you do not archive the finished film. You archive the project. The bins, the versions, the notes, the sources, the conform. The finished file answers exactly one question. The project answers all the others, including the one that always eventually gets asked, which is why did we do it that way.
I learned that hardest on stop-motion. I directed the Forno de Minas manifesto through WMcCann, built with Estudio Fue, and on that kind of job every frame is a decision made before the camera rolls and revision is close to impossible. When you cannot go back, the record of how you arrived becomes the actual asset. The deliverable is just the part the client sees.
Generative tools shipped with a chat log where the project file should be. A chat log is not an archive. It is a conversation stored in someone else's account, governed by someone else's roadmap, and now, in this case, ending on someone else's Sunday.
The archive discipline, borrowed from post
This is the usable part, and it takes minutes rather than an afternoon.
The master leaves the tool the same day it is made. Full resolution, straight into the job folder. Not the Downloads folder, which is not a folder, it is a landfill.
Write the prompt down next to the file. A plain text file with the same name. Prompt, model, model version, date, seed if the tool gives you one, and one line about what you were trying to get. That text file is smaller than a thumbnail and it is worth more than the image.
Name it for a stranger. Date first, then job, then shot, then version. You are not naming it for yourself this week. You are naming it for whoever opens the drive in four years, and there is a decent chance that person is you with no memory of any of this.
Two copies, two places, and one of them somewhere you do not work. A second cloud account inside the same workflow is not a backup, it is the same risk with a different logo.
Sort by what is derivable. Anything you can reproduce from the prompt is a cache. Anything you cannot is a master. Most people have this exactly backwards and are lovingly backing up their caches.
Once a quarter, try to rebuild one old image from your own notes. If you cannot, your record is incomplete and you have found out cheaply instead of expensively.
Two things make this more urgent than it looks. The first is that provenance metadata is fragile in exactly the way people assume it is not, which I went through when the watermark rules went live and the first export deleted them. The second is that owning the weights is becoming a real option rather than a slogan, which is the genuinely important part of the LTX 2.5 release. A model on your own disk cannot be retired by a blog post.
Ask for more of these, not fewer
The reflex response to August 30 will be to demand permanence. Keep the old tool. Promise us it will always be there.
Nobody can promise that and nobody should pretend to. Software gets retired. It has always been retired. The reasonable thing to ask for is not permanence, it is procedure: notice measured in weeks, an export path that returns full resolution, and a date written in public. OpenAI did all three here, which is why this is a boring story and should stay one.
The unreasonable thing is to keep treating a chat window as a filing cabinet and then act surprised when the building it was in gets sold.
Your images were never really in DALL-E. They were in an account, on a service, under terms you agreed to without reading, and that was always true of every tool any of us has ever rented, including the ones I use every day and will keep using tomorrow. That is not a betrayal. It is just the shape of the arrangement.
The part that belongs to you is the part you carried out of the building. It always was. August 30 is only the week somebody bothered to tell you.
Sources and verification
Facts verified August 24, 2026 against two or more independent sources per claim.
- The official DALL-E GPT is retired from ChatGPT on August 30, 2026, and OpenAI advises users to download images they want to keep: Tom's Guide, "You have until August 30 to save your ChatGPT DALL-E images"; Windows Report, "OpenAI will retire the official DALL.E GPT on August 30"; Crypto Briefing, "OpenAI retires DALL-E image generation from ChatGPT, pushing users to newer models" (published August 3, 2026); TechJournal, "OpenAI Is Retiring DALL-E in ChatGPT: What to Do" (published August 24, 2026); and the OpenAI Help Center ChatGPT release notes.
- Roughly thirty days of notice, replacement by ChatGPT Images running gpt-image-1 and gpt-image-1-mini, and user-created GPTs with image generation remaining functional: Crypto Briefing, corroborated on the replacement by TechJournal, which describes ChatGPT Images as a newer built-in generator with better instruction following, better text inside images, and editing of existing pictures.
- DALL-E 2 and DALL-E 3 model snapshots were removed from the API on May 12, 2026, with gpt-image-2, gpt-image-1 and gpt-image-1-mini given as recommended replacements: OpenAI developer documentation, deprecations page.
- No claim is made here about how many images any individual user has stored, about OpenAI's internal reasons for the consolidation, or about what any specific export path returns today, because none of that is published and export behaviour has to be checked per account. The comparison to silent model revisions is stated as an argument, not as a documented event.
