DALL-E 3 Left OpenAI's API on May 12: What Ad Makers Use Now

DALL-E 3 isn't running anywhere we could reach

DALL-E 3, OpenAI's text-to-image model, wasn't running on any surface we could reach on November 4, 2026. OpenAI removed it from its API on May 12, 2026, and retired ChatGPT's official DALL·E GPT on August 30. Bing Image Creator's FAQ had dropped it by the end of September. If you searched for it because you need an ad image today, you'll have to use another model, and most of this post is about which one and what it costs.

Of the top search results we could open, none gives an end date for each surface. The first still calls the API shutdown upcoming and says you can reach the model through OpenAI's API.

We read OpenAI's developer docs that same day. For its help center and main site, which wouldn't open for us, we relied on Internet Archive copies, each dated where we use it. The Novoads side comes from our own run on a model we sell. We didn't test DALL-E 3 for this post, and there's nowhere left we could reach to test it.

What DALL-E 3 was

What it made, and what it couldn't

OpenAI added DALL·E 3 to its API on November 6, 2023. A copy of its model page from April 4, 2026, says "DALL·E 3 currently supports the ability, given a prompt, to create a new image with a specific size." Text was the only input it took, and OpenAI's guide said "The Image Edits endpoint is not available for DALL·E 3," so it couldn't touch an image you already had.

It also changed what you typed. That guide, in a copy from January 28, 2026, says "OpenAI automatically rewrites your prompt for safety reasons and to add more detail." You couldn't turn that off, though the guide did offer a line you could add to your prompt to bring the result closer to what you'd asked for.

The same copy lists the model's limits in OpenAI's own words. It "struggles with rendering legible text," it "has trouble following precise instructions," and it's "not able to generate highly photorealistic images." In an ad, those three hit the headline, the brief and the product shot. So going by OpenAI's own description, we think it was a weak fit for ad work well before it ended.

How each page spells the name

You'll see the name written several ways. OpenAI writes DALL·E 3, with a middle dot, and its model ID is dall-e-3. Bing wrote DALL-E3 in its FAQ and DALL-E 3 on its help page. A Microsoft FAQ for Designer from 2024 wrote DALL.E 3, and what people type into search is dall-e 3.

Sizes, quality settings and prices while it ran

It came in three sizes and two quality settings, standard (the default) and hd, and OpenAI charged a flat price per image. The prices here come from a copy of OpenAI's pricing page dated February 7, 2026. The pricing page we read on November 4 has no DALL·E row at all. Treat them as history.

Quality	1024 x 1024	1024 x 1792	1792 x 1024
Standard	$0.04	$0.08	$0.08
HD	$0.08	$0.12	$0.12

OpenAI's own pages disagreed on the two non-square sizes. The pricing page and the guide said 1024 x 1792 and 1792 x 1024. The model page's April 4 copy showed the same prices against 1024x1536 and 1536x1024. We've gone with the pricing page and the guide, because they agree with each other.

Where DALL-E 3 ended, surface by surface

May 12 applies only to the API. Each surface went on its own schedule, and we couldn't find a date for two of them (Azure and Designer). Anything below without its own date was read on November 4, 2026.

Surface	What happened	When	What we read
OpenAI's API	Removed from the API	May 12, 2026	OpenAI's changelog and deprecations page
ChatGPT's official DALL·E GPT	Retired	August 30, 2026	OpenAI's help article, copies from August 27, September 21 and October 23, 2026
Bing Image Creator	Gone from its FAQ's model list	By the end of September 2026, and no page we read names the day	FAQ copies from August 22 and September 30, 2026, and the FAQ as it stands now
Azure (Microsoft Foundry)	No longer in Azure's model list	No date in the Azure docs we read	Azure's list of models it sells directly, updated October 29, 2026
Microsoft Designer	Not stated, because its FAQ names no image model	No date	Designer's FAQ, updated September 11, 2026

OpenAI's API, May 12, 2026

OpenAI says it told developers on November 14, 2025. Its changelog says "DALL·E model snapshots dall-e-2 and dall-e-3 were deprecated and removed from the API on May 12, 2026." The deprecations page gives the same date with no extension, and the model page now reads "Status: Shut down." OpenAI uses "shut down" to mean "a model or endpoint is no longer accessible."

Developers' reports fit that date. A LiteLLM pull request from May 13, 2026, says the API was answering "The model 'dall-e-3' does not exist," and a GitHub issue reports the same error on direct requests sent on October 25. OpenAI gives no hour, so we won't either. We didn't try an API call of our own, so we can't tell you it stopped for every account everywhere.

OpenAI's deprecations page also says "In some cases, developers may be able to provision dedicated capacity for continued access after a model's shutdown date." That sentence is general and doesn't name DALL·E, but it's why we say "removed from OpenAI's API" and leave it there.

ChatGPT's official DALL·E GPT, August 30, 2026

ChatGPT's date came later. A copy of OpenAI's help article from August 27, 2026, says "On August 30, 2026, we're retiring the official DALL·E GPT from ChatGPT," and it asks you to download any images you want to keep before then. Copies from September 21 and October 23 answer the same question with "No - we've retired the official DALL·E GPT from ChatGPT," and they send you to ChatGPT Images.

The August copy adds "User-created GPTs with image generation enabled are not affected." It doesn't say which model those GPTs run, and the two later copies don't carry the line. So we wouldn't read it as DALL-E 3 surviving somewhere inside ChatGPT. If you meant to keep images from the official GPT, that deadline has passed, and we didn't check what's left in anyone's account.

Bing Image Creator, by the end of September 2026

Going by its FAQ, Bing kept DALL-E 3 after OpenAI's API dropped it. Bing's blog offered "the choice of two models," GPT-4o and DALL-E3, from August 6, 2025. Its FAQ still said "If you select DALL-E3" in a copy from August 22, 2026. The copy from September 30 lists only MAI-Image-2.5-Flash and MAI-Image-2e, and so does the current FAQ. No page we read names the day, so we can only say by the end of September.

Two things on Bing's site could mislead you. Its Mode menu offers a setting called "Vivid and natural," which belongs to MAI-Image-2.5-Flash. It sounds like DALL-E's old style setting, which OpenAI's API reference now calls a "Legacy style parameter for retired image models," but it's a different thing. The page also says "Sign in to create more and unlock additional models." We read Bing signed out, from Spain, so we can't speak for a signed-in account or another market.

Azure's model list and Microsoft Designer

Azure's list of the models it sells directly, updated October 29, 2026, names six GPT image models and no DALL-E. The Azure docs we read don't name dall-e-3 or give it a date, so we'll leave it at "no longer in Azure's model list."

Designer is the murkiest of the five. A Microsoft FAQ from 2024 says Designer makes new images "with DALL.E 3 and GPT 3.5 integrated into the design experience from the start." Designer's current FAQ, updated September 11, 2026, names no image model, so we can't tell you whether it runs DALL-E 3 now. For ad work, the license line matters more, and that FAQ says "Designer is licensed for personal, non-commercial use with your Microsoft account."

Copilot isn't in the table, because we didn't read a Microsoft page that says which image model it runs. Designer's FAQ describes making an image through the Copilot icon in Word and PowerPoint, and it doesn't name a model there either.

Lines you may still run into

• OpenAI's own DALL·E 3 post still says "DALL·E 3 is now available to all ChatGPT users, as well as to developers through our API." In a copy from November 1, 2026, that line sits under a banner reading "This post introduced DALL·E 3," so it's a launch post kept for the record.
• The DALL·E 3 model page marks its image generation endpoint "Supported" and also says "Status: Shut down." We read that row as old listing data, because the sentence under the status line says the model is gone from the API.
• The first search result, an explainer dated February 13, 2026, says the model "was removed from ChatGPT without warning in December 2025." OpenAI's help article has the official DALL·E GPT running until August 30, 2026. By that article's August copy, ChatGPT's main image maker was already ChatGPT Images 2.0, so the explainer may mean an earlier change of default, but we didn't read a dated OpenAI page for one. The explainer also names GPT Image 1.5 as the replacement, and OpenAI's deprecations page now has that model leaving on December 1, 2026.
• A post on OpenAI's developer forum from April 9, 2026, "Deprecation Reminder: DALL·E will be shut down on May 12, 2026," gets the date right. Its author shows no staff label, so we don't read it as OpenAI speaking. And the two replacements it suggests, gpt-image-1.5 and gpt-image-1-mini, are both on the December 1 list.

What OpenAI points to now

GPT Image 2.5 for new work

OpenAI's answer, on the pages we read November 4, 2026, is GPT Image 2.5. The DALL·E 3 model page says "We recommend GPT-Image-2.5 Sunburst for current image generation and editing," and the image guide says "For new integrations, use one of the GPT Image 2.5 models described above." OpenAI released the two models on September 8, 2026. Its guide says to choose "Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation." We've written up how that choice plays out for ads in our GPT Image 2.5 post.

Where that leaves GPT Image 2

GPT Image 2 is harder to place, because OpenAI's pages disagree. Its model page calls it "our state-of-the-art image generation model." The image guide files it under a heading that reads "Earlier GPT Image models" and says the details there "apply to earlier models, not Sunburst or Flare." The deprecations page posts no shutdown date for it, and it shows up there only as a replacement. OpenAI says that, unless safety or compliance concerns require a faster timeline, generally available models get at least 6 months' notice before retirement. That's the warning we'd expect if it's ever retired.

The May replacement list is out of date

The May entry that removed DALL-E 3 recommends "gpt-image-2, gpt-image-1, or gpt-image-1-mini," and the deprecations page repeats that list in its DALL·E rows. If you're moving something off DALL-E 3 now, don't work from it. Further up, the same page has these rows.

Model	What the deprecations page says	Replacement it names
gpt-image-1	Shut down on October 23, 2026	gpt-image-2.5-sunburst or gpt-image-2.5-flare
gpt-image-1-mini	Leaving the API on December 1, 2026	gpt-image-2.5-sunburst or gpt-image-2.5-flare
gpt-image-1.5	Leaving the API on December 1, 2026	gpt-image-2.5-sunburst or gpt-image-2.5-flare
chatgpt-image-latest	Leaving the API on December 1, 2026	gpt-image-2.5-sunburst or gpt-image-2.5-flare
gpt-image-2	No shutdown date posted	Appears only as a replacement itself

As we read those rows, GPT Image 2 is the only one of the May three without a date. OpenAI doesn't put it that way.

ChatGPT and Codex run different things

Inside OpenAI's own apps, the model depends on which one you're in. The ChatGPT app has ChatGPT Images, and the October 23, 2026, copy of the help article says "ChatGPT Images 2.5 brings sharper details, more precise editing, and faster image generation" and "ChatGPT Images is available on all tiers." ChatGPT's Learn page on Codex says "Built-in image generation uses gpt-image-2 and counts toward your general Codex usage limits."

What changes if you used DALL-E 3 for ads

The DALL-E 3 column shows the model as OpenAI described it, and the GPT Image 2 column shows OpenAI's API as we read it.

What changes	DALL-E 3 while it ran	GPT Image 2 on OpenAI's API
Input	Text only	Text and images
Editing an image	Not available	Supported
Sizes	Three fixed sizes	Any width and height divisible by 16, with an aspect ratio between 1:3 and 3:1, up to 3840x2160
Quality settings	standard and hd	low, medium and high
Your prompt	Rewritten every time, with no off switch	The Image API returns no revised prompt, and the Responses API's image tool revises it first
Output	A URL by default	Always base64
Text in the image	"struggles with rendering legible text"	Significantly improved, but it "can still struggle with precise text placement and clarity"
Price	Flat per image	Per token, $30.00 per 1M image output tokens

The text row is the one we'd look at first, because a headline and a label have to read correctly in an ad. The prompt row comes next. DALL-E 3 rewrote whatever you typed, and OpenAI's Image API doesn't hand you back a rewritten prompt for GPT image models. So we think the brief you type has to do more of the work. Our guide to GPT Image 2 prompts for ads covers how to write one.

If you've got code still calling dall-e-3, plan for more than a new model name. Its default output was a URL to a hosted image, and OpenAI's reference says GPT image models "always return base64-encoded images."

What ad makers can use now, in Novoads

We can't tell you what ad makers at large are running, because nothing we read counts it. But we can tell you what we sell and where we'd start. Novoads has never sold DALL-E 3. We sell GPT Image 2, which OpenAI released on April 21, 2026, and both GPT Image 2.5 models. Each costs a whole number of credits, and that number doesn't change with the image's shape.

Model in Novoads	Low	Medium	High
GPT Image 2	1 credit	1 credit	4 credits
GPT Image 2.5 Sunburst	1 credit	1 credit	2 credits
GPT Image 2.5 Flare	1 credit	1 credit	2 credits

For a new ad set, we'd start with a GPT Image 2.5 model, because that's where OpenAI points new work and because at high it costs half the credits GPT Image 2 does. That call rests on OpenAI's direction and our own price list, since we didn't run the 2.5 pair for this post.

One ad prompt on GPT Image 2

GPT Image 2 is the one we can show you. On November 3, 2026, we tried a single prompt for a cold brew ad in Novoads. We made four images in all, low and medium at 2:3, then low and medium at 1:1. The prompt asks for a can of Tidewell sitting on wet granite beside a sunny kitchen window, spells out the label and asks for a headline once. The strip shows the two 2:3 takes. It's a GPT Image 2 render, and we've got no DALL-E 3 image to set beside it.

Made in Novoads with GPT Image 2 · one cold brew can ad prompt at 2:3, low on the left and medium on the right · 1 credit each · rendered 2026-11-03

The headline, "Cold brew, zero wait," came out clean on all four takes. The low take on the left garbled the can's label, while the low take at 1:1 (not in the strip) spelled it right. Medium spelled TIDEWELL right both times. That's two takes per quality, so treat it as what happened in that run rather than as a rate.

We'd also planned a high take, which costs 4 credits. It failed on all five tries that day, and the credits came back each time. So we've got no high image to show you. Our GPT Image 2 vs Nano Banana Pro comparison runs the same three prompts on both models, if you want to see it against another maker's model.

What one image costs, against OpenAI's printed figure

Prices as of November 4, 2026. You buy Novoads credits through a plan, and you pay from the first month. Starter is $15 a month for 200 credits, monthly only. That's $0.075 a credit. Plus is $49 a month for 1,000 credits. That's $0.049 a credit. All the plans are on our pricing page.

Screenshot: the Novoads pricing page with Monthly selected, taken 2026-11-04

This table puts DALL-E 3's old prices, from the February 7 copy above, beside OpenAI's printed figure for GPT Image 2 and the same image in Novoads.

Image and tier	DALL-E 3 while it ran	GPT Image 2 on OpenAI's API	GPT Image 2 in Novoads, Starter	GPT Image 2 in Novoads, Plus
Square, low	No low tier	$0.006	$0.075	$0.049
Square, medium (DALL-E 3 Standard)	$0.04	$0.053	$0.075	$0.049
Square, high (DALL-E 3 HD)	$0.08	$0.211	$0.30	$0.196
Tall, low	No low tier	$0.005	$0.075	$0.049
Tall, medium (DALL-E 3 Standard)	$0.08	$0.041	$0.075	$0.049
Tall, high (DALL-E 3 HD)	$0.12	$0.165	$0.30	$0.196
One square plus one tall, medium	$0.12	$0.094	$0.15	$0.098

A square is 1024 x 1024 in both OpenAI columns. Tall means 1024 x 1792 for DALL-E 3 and 1024 x 1536 for GPT Image 2, and it's 2:3 in Novoads. The rows line up tiers and don't promise equal quality, because DALL-E 3's standard and hd don't map onto low, medium and high. Our run's medium takes also came back larger than OpenAI's printed sizes, at 1536 x 2304 and 2048 x 2048, so the pixels don't line up exactly either.

That table cuts both ways.

• On Plus, a medium square comes in about 8% under OpenAI's printed figure, and a high square about 7% under.
• At 2:3, OpenAI's figure is lower, and Plus costs about 1.2 times as much.
• Starter costs about 1.4 times OpenAI's figure on a square and about 1.8 times at 2:3.
• At low, OpenAI is far cheaper, with Plus at about 8 to 10 times its figure, because we charge the same credit for low and medium.

OpenAI's figures come with limits too. They cover image output only, and OpenAI says "You should still account for text and image input tokens when estimating the total cost of a request." Text tokens cost $5.00 per 1M, and image input tokens cost $8.00 per 1M. You'd also be building on its API and paying per token, though its Batch tier takes 50% off. If you're a developer making lots of low-quality images, OpenAI's API is the cheaper route, and we'd tell you to use it. Our GPT Image 2 pricing post has the rest of the math by shape.

On Starter, your credits get you 200 images a month at low or medium, or 50 at high. On Plus, it's 1,000 or 250.

What each page says about using the image

OpenAI's Terms of Use, effective January 1, 2026, and read in a copy from November 2, 2026, say that "to the extent permitted by applicable law," you "own the Output." They also warn that "output may not be unique and other users may receive similar output from our Services." We didn't read OpenAI's Europe terms. If you still have old DALL-E 3 images, OpenAI's launch post says they're "yours to use and you don't need our permission to reprint, sell or merchandise them."

On provenance, a copy of OpenAI's help article from October 22, 2026, says "Images generated with ChatGPT, Codex, or the OpenAI API include both signals," meaning Content Credentials and SynthID watermarks. It calls SynthID "an invisible watermarking technology," and it adds "Coverage may vary by product, model, export path, file type, and when the content was created." The exceptions it names include gpt-image-1 and gpt-image-1.5 and leave out gpt-image-2, so we read GPT Image 2 on OpenAI's own surfaces as covered. OpenAI doesn't say that outright, and we didn't check the files from our run for those signals.

Our own FAQ says "All videos generated with Novoads are fully licensed for commercial use." That line names videos, so we won't stretch it past them here. Bing Image Creator's terms say "The Online Services are for entertainment purposes," and with Designer's license line above, we wouldn't build a paid ad on either one.

Which route fits your next image

• If you just want to try ChatGPT's image maker, the October copy of the help article says it's on all tiers, so that's the shortest way in.
• If you'd like a no-cost image from another maker, Google's page lists Nano Banana as "Available for free," and our post on whether Nano Banana is free goes through what that covers.
• If you're a developer with code to move, go straight to a GPT Image 2.5 model on OpenAI's API, as OpenAI asks, and leave the May list alone.

Read the terms of whichever you pick before you run anything paid.

If you're making ads, we built Novoads for that job. You pick GPT Image 2 or a GPT Image 2.5 model, and the same credits also buy video ads, product photos, talking actors and voiceovers, so your images and videos draw on one balance. Nano Banana Pro is there too, if you want a second maker's model to test against.

The next date on OpenAI's list is December 1, 2026, when gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest leave the API. If anything of yours still runs on one of those, that's the one to move next.