A florist in a small town needs a picture of a bouquet she has not made yet. It is for next month's seasonal promotion, the flowers are not in season, and the shoot would cost more than the promotion is likely to earn. So for three years she has done what most small businesses do, which is use a stock photo of somebody else's bouquet and hope nobody notices it does not match what arrives.
Yesterday, the tool that solves that problem quietly became free.
OpenAI retired the DALL-E GPT from ChatGPT on 30 August 2026, closing the last visible DALL-E surface inside the product. The announcement came on 31 July, giving people about thirty days to save their work. What replaced it is not a downgrade dressed up as a migration. It is a considerably better tool that costs nothing on the free tier, and a lot of small businesses have not registered that it happened.
What actually happened
The retirement was the end of a long wind-down rather than a sudden decision. OpenAI removed the dall-e-2 and dall-e-3 models from its API on 12 May 2026, having announced that move back in November 2025. The DALL-E GPT inside ChatGPT survived a few months longer as the last place the name still appeared, and on 30 August it went too.
Image generation did not go anywhere. It moved to ChatGPT Images, which runs on gpt-image-2 and has been live since 21 April 2026. If you have generated an image in ChatGPT at any point since the spring, you were most likely already using the new system without noticing, because OpenAI made it the default long before it removed the old one. Custom GPTs that generate images are unaffected.
The one genuinely time-sensitive thing was saving old work. Images you created through the DALL-E GPT needed downloading before 30 August. If you missed that window and had something you cared about in there, that is the part of this story with an actual cost, and it is worth checking whether anything of yours was sitting in that history.
What you get for free
The headline for a small business is that ChatGPT Images is the default image tool across every account tier, free included. Free accounts get roughly two to three images per rolling 24 hour window, and the detail worth knowing is that the window resets 24 hours after your first generation rather than at midnight. That catches people out: if you generate at 9pm on Monday, your quota returns at 9pm on Tuesday, not when the date changes.
Output reaches 2048 pixels, with aspect ratios ranging from 3:1 through to 1:3, which covers a wide banner, a square social post, and a vertical story format without cropping something that was never composed for it. A single prompt can return up to eight images that stay consistent with each other, which matters more than it sounds: consistency across a set is the thing that has historically made AI images unusable for anything beyond a one-off.
Commercial rights are included, and they are included on the free tier too. OpenAI's terms allow commercial use of every output. You own what comes back and you can put it on a product page, an advert, or a menu without a licence question hanging over it. For a business that has been paying for stock photography subscriptions to avoid exactly that question, this is the part of the change with real money attached.
Why it is genuinely better
It would be easy to treat this as a rebrand, and it is not. The improvements are in the specific places where DALL-E used to fail in ways that made its output unusable for business purposes.
Prompt adherence is the big one. DALL-E had a habit of producing something beautiful that was not what you asked for, which is charming in a toy and useless in a tool. If you specified a product on a marble surface lit from the left with a copper vase behind it, you got roughly two of those three things and a negotiation about the rest. The newer model handles compositional requests far more reliably, including spatial relationships between objects and specific lighting conditions, which is precisely the vocabulary a business uses when describing what it actually needs.
Editing is the second. Where the old workflow meant regenerating from scratch and hoping the next attempt kept the parts you liked, the current one supports targeted edits to an existing image. That changes the shape of the work from gambling to iterating. Generation is also up to four times faster, and fewer visual artefacts survive into the output, which reduces the number of images you discard for having a strange extra finger or a piece of text that dissolves into nonsense.
Text rendering deserves a specific mention because it was the single most reliable way to spot AI imagery. Legible words inside a generated image used to be close to impossible. It is now workable, which opens up the category of asset a small business most often needs, meaning something with a price, a date, or a name actually on it.
The watermark nobody mentions
Here is the part that did not make the headlines and matters more than most of what did.
Every image gpt-image-2 produces carries two marks. The first is C2PA metadata, an industry provenance standard that records the image as AI-generated. The second is an imperceptible pixel-level watermark embedded in the image data itself. Both are applied on every tier, free and paid alike, and neither is optional.
Their durability differs in a way worth understanding. The C2PA metadata is fragile and strips out if someone screenshots the image, because a screenshot is a new file that never carried the original metadata. The pixel watermark is the resilient one and survives most common editing operations. So the practical position is that an image you generated, cropped, and posted is still identifiable as AI-generated by anyone who checks, even though nothing about it looks different to a human eye.
To be clear, neither mark restricts your commercial use. OpenAI's terms are unchanged and you still own the output. What the watermark affects is disclosure, not licensing, and those are different questions that are easy to conflate. This is the same mechanism and the same distinction we went through in the Claude text watermark article, applied to pictures instead of prose.
Where this fits in a small business
The disclosure question is where this stops being a fun free upgrade and becomes something you need a position on, because two separate sets of rules now care about it.
If you sell into the EU, Article 50 of the EU AI Act requires AI-generated content to be clearly labelled and to carry machine-readable marks, with the deadline for synthetic image content landing on 2 December 2026. The machine-readable half is already handled for you by the C2PA metadata, which is a genuine convenience. The visible labelling half is your responsibility and has a date attached. We covered the full timeline in the EU AI Act deadline guide.
If you sell on a marketplace, the platform rules bite sooner and harder than the regulation does. Etsy in particular now requires AI disclosure on listings and matches what sellers declare against its own image analysis, which means a watermarked image and an undeclared listing is a combination the platform can detect directly. Marketplace enforcement tends to arrive as a removed listing rather than a warning letter, which makes it the more immediate risk for most sellers.
For everything outside those two contexts, meaning your own website, your own social accounts, your own printed materials, there is no rule requiring you to announce it. The honest recommendation is still to think about it once rather than never. An AI-generated photo of a product that looks nothing like what actually arrives is a customer service problem regardless of whether any regulator or platform ever notices.
The honest limits
Two or three images a day is a real constraint on the free tier, and it is the constraint that will decide whether this tool is useful to you. Anyone who has generated images knows the first attempt is rarely the one you use. A quota of three attempts means one usable asset on a good day and nothing on a bad one, which is fine for occasional needs and useless for producing a week of social content.
Consistency across separate sessions remains genuinely unsolved. Eight images from one prompt hold together well. Eight images generated across three days do not, and the same product described identically will come back with a different material quality, a different light, a different sense of scale. For a set of product variants meant to sit together on a page, that variation is exactly what a customer notices without being able to name it.
And the thing worth saying plainly at the end: the constraint on your marketing was never the picture. It was knowing what to say and to whom. A free image generator removes a cost that was real and does not touch the part that was actually hard. The florist still needs to know that her seasonal promotion should lead with the arrangement people buy for their mothers, not the one she is proudest of. No model has ever been able to tell her that.