DALL-E AI Image Generator in 2026: What Replaced It and How to Use It Now

If you searched for the DALL-E AI image generator, here’s the thing most of the pages ranking for it won’t say: OpenAI retired it. DALL-E 3 left ChatGPT in March 2025. Both dall-e-2 and dall-e-3 were switched off in the API on 12 May 2026. Bing Image Creator, the old “free DALL-E” route, now runs Microsoft’s own models.

The name stuck around anyway. People still type it, and a crowd of look-alike sites still use it. So this page does the job the old version of it should have done: it tells you what the DALL-E name points to in October 2026, which tool to open today, what each option actually costs, and how to move old DALL-E API code to the successor without breaking anything.

I’m not writing this from a review desk. My business runs on autonomous agents, and every image on this site comes out of an unattended pipeline that has logged 626 generated images since 31 January 2026. Not one of them was made with DALL-E. I’ll show you why, and what the pipeline looks like.

Freshness note: everything marked READ below was read off the vendor’s own page on 14 October 2026. Prices and limits in this category move every few months, so check the date before you budget off it.

Is the DALL-E AI Image Generator Still Available in 2026?

Short answer: not from OpenAI, and not under that name. Here’s the timeline, each step taken from a primary source:

Date	What happened	Source
Jan 2021	The first DALL-E is announced	OpenAI / Wikipedia
Apr 2022	DALL-E 2 announced, opened to everyone in Sep 2022	OpenAI
Oct 2023	DALL-E 3 ships in ChatGPT and Bing Image Creator, with API access in early November	OpenAI / Microsoft
Mar 2025	ChatGPT replaces DALL-E 3 with native GPT Image generation	Wikipedia, citing TechCrunch
14 Nov 2025	OpenAI tells developers both DALL-E API models will be removed	OpenAI deprecations page
12 May 2026	dall-e-2 and dall-e-3 shut down in the API	OpenAI deprecations page
8 Sep 2026	ChatGPT Images 2.5 launches, with GPT-Image-2.5 Flare and Sunburst in the API	OpenAI

Sources: OpenAI’s API deprecations page (the 2025-11-14 DALL·E entry), the ChatGPT Images 2.5 announcement, and Wikipedia’s DALL-E history.

OpenAI’s model page for DALL-E 3 now has a red “Deprecated” badge and says plainly: “DALL·E 3 has been deprecated and removed from the API.” The original DALL-E 3 launch post is still up, but it now carries a banner sending readers to ChatGPT Images 2.5 for “the current experience.”

So when someone says “use DALL-E” in 2026, they almost always mean one of three things:

• “Make me an image in ChatGPT.” That’s ChatGPT Images now, not DALL-E.
• “Use the free one on Bing.” That’s Microsoft’s MAI-Image models now, not DALL-E.
• “Call OpenAI’s image API.” That’s the GPT Image model family now. A request for dall-e-3 no longer works.

None of this makes the old DALL-E knowledge worthless. The prompting habits carry over almost one-for-one, and so does the idea of a text-to-image call sitting inside a workflow. What changed is the model name, the price shape, and a few API parameters. That’s the stuff that quietly breaks automations, and it’s what the rest of this page covers.

What Replaced the DALL-E AI Image Generator (and Where Each Version Lives)

The DALL-E name used to cover one model reached through three doors. Today each door leads to a different model, and they’re priced differently.

1. ChatGPT Images 2.5 (the consumer replacement)

OpenAI announced ChatGPT Images 2.5 on 8 September 2026, saying it’s “available to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web.” (READ). It does the conversational thing DALL-E 3 was loved for: describe the picture, then refine it in plain English (“warmer light”, “move the mug left”). OpenAI says it keeps adding C2PA metadata and invisible watermarking to images made with its tools.

I’m not quoting per-plan image limits here. OpenAI didn’t publish them on the announcement page, and I won’t guess.

2. GPT-Image-2.5 Flare and Sunburst (the API replacement)

For developers, the image generation guide now names two current models (READ):

• gpt-image-2.5-flare: OpenAI’s “default choice for most applications”. The launch post says it produces higher-quality images than GPT-Image-2 at 50% lower latency.
• gpt-image-2.5-sunburst: for “premium visual workflows” where precise edits matter. It’s slower and more exact.

The deprecations notice for DALL-E listed gpt-image-2, gpt-image-1 and gpt-image-1-mini as replacements back in November 2025. The DALL-E 3 model page has since moved on and now points to Sunburst. If you’re migrating today, start with Flare.

3. Bing Image Creator (the free route, no longer DALL-E)

Bing Image Creator is still free with a Microsoft account, but its help page now offers two Microsoft models: MAI-Image-2.5-Flash and MAI-Image-2e (READ). Uploading an image to edit forces MAI-Image-2.5-Flash. You get 15 fast creations a day, then a slower standard speed, capped at 200 prompts per 24 hours. Plenty of older guides still call it “free DALL-E 3 on Bing”. That’s no longer true.

Old route	What it is in Oct 2026	Best for
DALL-E 3 in ChatGPT	ChatGPT Images 2.5	One-off images, conversational edits
DALL-E 3 on Bing	Bing Image Creator on MAI-Image-2.5-Flash / MAI-Image-2e	Free experimenting with a Microsoft account
dall-e-3 in the API	gpt-image-2.5-flare (or -sunburst)	Apps, automations, agents

Those “DALL-E” Sites in Google Are Not OpenAI

I pulled the live US organic results for “dall e ai image generator” on 14 October 2026. Look at what fills the top seven:

• #1 dall-efree.com: its page title reads “Powered by Dall-E”, but its own FAQ calls it “an independent AI image generation platform” and shows logos for nine model vendors (Black Forest Labs, ByteDance, Google, Ideogram, Kling, OpenAI, Qwen, Runway and others). Plans run from a free tier with 2 credits up to $50 a month (READ). It’s a multi-model credit reseller using the DALL-E name.
• #2 and #6 openai.com: the 2023 DALL-E 3 and 2022 DALL-E 2 launch posts. Both are history pages now.
• #3 Wikipedia: accurate history, including the March 2025 replacement in ChatGPT.
• #4 NightCafe: a multi-model credit platform with a DALL-E 3 landing page.
• #7 dalle3.ai: a 2023 affiliate page that, as of today, still says “DALL-E 3 is not available yet” and sends every click to NightCafe.

None of this is a scam accusation. Reselling access to many models is a legitimate business. But if you’re choosing a tool, know what you’re buying. With a wrapper you pay its credit price for somebody else’s model, and you don’t always know which model ran. And if you’re building something that has to keep working, you want the model name in your code, not hidden behind a reseller’s dropdown.

Here’s my rule of thumb: if a site has “DALL-E” in its domain and isn’t openai.com, assume it isn’t DALL-E. The real model can’t be called anymore, so nobody can sell it to you.

How to Use the DALL-E AI Image Generator’s Successor Today

Pick the route that matches what you’re doing, not the one with the most famous name.

If you need a few images this week: ChatGPT

1. Open ChatGPT and say you want an image. Describe subject, setting, style and light in one or two sentences.
2. Look at the first result before you add detail. Most people over-specify up front, then fight the model.
3. Refine in plain language, one change per turn: “make the background darker”, “same scene, from above”.
4. When it’s right, download it. If you need an exact size, say so (“16:9 landscape”) instead of cropping later.

If you want free and don’t mind waiting: Bing Image Creator

Sign in with a Microsoft account, pick MAI-Image-2.5-Flash or MAI-Image-2e, and type your prompt. After the 15 daily fast creations you drop to standard speed, so batch your experiments early in the day. It’s great for trying out a visual idea. I wouldn’t build a business process on it, because there’s no API, no SLA, and the model under the button has already changed once.

If an image has to appear without you: the API

This is the only route that belongs in an automation. A minimal call to the current model looks like this (shape taken from OpenAI’s image generation guide):

curl -s https://api.openai.com/v1/images/generations \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.5-flare",
    "prompt": "flat geometric illustration of a delivery van at dawn, teal and navy, no text",
    "size": "1536x1024",
    "quality": "medium"
  }' | jq -r '.data[0].b64_json' | base64 --decode > van.png

Three settings decide your cost and speed: size (recommended sizes are 1024×1024, 1536×1024 and 1024×1536, and custom sizes like 1536×864 work if both edges are multiples of 16), quality (low through max on the 2.5 models) and how many images you ask for. OpenAI’s own advice is to draft at low and only compare higher settings for final assets. Follow it. Most blog and social images never need more than medium.

A generated file isn’t a published image yet, though. In my stack the API call is one step out of six, and the other five (compression, CDN, upload, featured-image assignment, logging) are where things actually break. More on that below.

Migrating Old DALL-E API Code Without Breaking Your Pipeline

If something you built still sends "model": "dall-e-3", it has been failing since 12 May 2026. The dangerous part is that a lot of automations fail quietly: a no-code scenario that swallows the error, a fallback that posts without an image, a blog that publishes with a blank featured slot. I wrote about this pattern in what happens to your automation when the AI model gets retired. The model you pinned by name is the part that gets switched off on someone else’s calendar.

Here’s the migration checklist I’d run:

1. Find every pinned model ID. grep -rn "dall-e" . across your scripts, workflow exports and config. Don’t forget environment variables and no-code templates.
2. Swap the model name to gpt-image-2.5-flare (or -sunburst if edit precision matters more than speed).
3. Fix the sizes. DALL-E 3 used 1792×1024 and 1024×1792 for wide and tall. Those aren’t in the new recommended list. Use 1536×1024 / 1024×1536, or a custom size that fits the multiples-of-16 rule.
4. Remove DALL-E-only options. Parameters like style: "vivid" and quality: "hd" belonged to DALL-E 3. Quality on the new models is low/medium/high/xhigh/max/auto.
5. Check how you read the result. OpenAI’s current examples decode a base64 field (b64_json). If your old code downloaded a returned URL, that step changes.
6. Re-budget. DALL-E was priced per image. GPT Image models are priced per token (next section). Run ten real requests, read the usage block, and work out your actual per-image cost from that, not from a blog post. This one included.
7. Make failure loud. If image generation fails, the job should stop and alert you, not publish without an image. My pipeline treats a missing image as a hard stop.

If your workflow lives in n8n or Zapier, the same list applies: open every node that calls OpenAI images and check the model dropdown. A retired model doesn’t always show up as an error in a visual builder until something downstream comes out empty.

Prompting That Still Works (Lessons From 626 Pipeline Images)

Good news: the prompting skill you built on DALL-E transfers to every current model I’ve used. The structure that works is boring and reliable:

• Subject first. One clear main thing. “A delivery van at dawn” beats “a scene showing logistics and growth”.
• Style second. Name it concretely: “flat geometric illustration”, “editorial photograph”, “isometric 3D render”.
• Light and palette third. “Soft morning light from the left, teal and navy palette.” This is where brand consistency lives.
• Composition last. “Wide shot, subject on the right third, empty space on the left for a headline.”

And a few lessons I only learned by running this unattended at volume:

Every image prompt on this site ends with “no text, no letters, no numbers, no logos.” Current models render text far better than DALL-E 2 did, and OpenAI’s own docs still list precise text placement as a limitation. One word in a prompt can trigger garbage lettering. When a prompt of mine said “labeled”, the model painted pseudo-text onto the image and I had to re-render it. So I never ask the image model for words. Headlines go on in HTML, where they’re crisp, searchable and editable.

Put the brand in a file, not in your head. My image style is one sentence stored in a brand config: dark background, teal and cyan accents, clean geometric lines, no text, no logos. Every image prompt the agents write appends it. That one line is why seven images made at midnight look like they belong together. (I made the same point about brand files in the brand kit generator guide: the file is the brand kit.)

Eyeball the first one, then sample. I look at the featured image of every post, plus any image whose prompt mentions charts, diagrams or labels. Those are the prompts that go wrong.

Specify the aspect ratio per destination. 16:9 for blog and Facebook, 1:1 for Instagram, 2:3 for Pinterest, 9:16 for vertical video. Cropping after the fact cuts off the subject far more often than you’d expect.

What I Run Instead of the DALL-E AI Image Generator: A Pipeline With Receipts

Here’s the part you won’t find in the top 10. The images on this blog, including the ones on this page, are made by a script my agents call with no human involved. It runs six steps:

1. Generate on Replicate (default model: flux-2-pro, JPEG out). See my Replicate breakdown for why I rent models by the run.
2. Compress through Tinify. This is a mandatory gate: if compression fails, the run stops. No uncompressed image reaches the site.
3. Upload to Cloudinary as a permanent JPEG. Social platforms get this URL, because WordPress may convert uploads to WebP and Meta’s APIs reject WebP. The full story is in my Cloudinary write-up.
4. Upload to WordPress media.
5. Optionally set it as the featured image of a post.
6. Log it to an Airtable ledger.

The ledger is the receipt. Here’s what it held when I queried it for this post:

Model (as logged)	Images
flux-2-pro (four label spellings combined)	518
flux-dev multi-LoRA (early experiments)	86
flux-pro (legacy default)	13
flux-dev	3
openai/gpt-image-1.5	2
ideogram-v3-quality	2
nano-banana-pro	2
DALL-E (any version)	0
Total, 31 Jan to 13 Oct 2026	626

Two honest notes. First, the ledger is messy: the same model was logged under four slightly different names, and nine rows record a day when Tinify rate-limited us and images were compressed locally instead. Logs written by agents drift unless you check them. Second, the two GPT Image rows were experiments. I have no head-to-head benchmark to sell you, so I won’t claim flux beats OpenAI’s models. My default is flux-2-pro because the script’s own dated comment says the cheaper model “produces weak output for blog use”, and so far nothing has forced a change.

Today’s run, for this exact page: 9 renders for 8 images. Tinify took the 8 keepers from 1334 KB to 412 KB combined (69% smaller, between 64% and 80% per file). The 9th render was a reject: the first version of the “how to use” image had faint pseudo-letters hidden in its light streaks, so I re-rendered it with a stricter no-glyphs prompt. That’s the eyeball check doing its job, and it’s why I never let a pipeline publish an image nobody has looked at at least once.

The lesson from all of this has very little to do with DALL-E. The generator is the replaceable part. The value sits in the pipeline around it: compression, CDN, format rules, logging, and a hard stop when a step fails. If DALL-E’s retirement broke your workflow, the fix isn’t finding the next famous model. It’s making the model one swappable line, the same way I treat every model in an AI social media system or a content creation pipeline.

DALL-E AI Image Generator Pricing in 2026 (What the Replacements Cost)

DALL-E pricing is gone along with DALL-E. Any page still quoting per-image DALL-E 3 prices, or “credits” from OpenAI, is describing a product you can’t buy. Here’s what you actually pay today:

Option	Price shape	Numbers (READ 14 Oct 2026)
ChatGPT Images 2.5	Included in ChatGPT plans	Available to all ChatGPT users per OpenAI. Per-plan image limits not published on the launch page
Bing Image Creator	Free with Microsoft account	15 fast creations/day, then standard speed, 200 prompts per 24h
GPT-Image-2.5 Flare / Sunburst (API)	Per token	$30 per 1M image output tokens, $8 per 1M image input, $5 per 1M text input. Batch is half price
flux-2-pro on Replicate (what I run)	Per run + per megapixel	$0.015 per run plus $0.015 per output megapixel on the model page. My script budgets ~$0.08 per image to stay conservative
Wrapper sites (e.g. dall-efree.com)	Monthly credits	Free (2 credits), $15, $30, $50 per month

API figures are from OpenAI’s pricing page. The important shift is in the API row. With per-token billing, two images at the same nominal size can cost different amounts depending on the quality setting and the model. OpenAI says this directly: equal token rates don’t mean equal cost per image. So don’t copy a per-image figure from anyone. Run your real prompts at your real settings and read the usage field. Streaming partial previews also cost extra (100 image output tokens each), which is easy to miss.

For a small business, a sensible decision path looks like this:

• Fewer than ~20 images a month, made by you: ChatGPT or Bing. Don’t build anything.
• Images that need to appear on a schedule (blog, social, product pages): an API plus a pipeline. The model bill will be the smallest line in it.
• Brand consistency matters more than novelty: pin your style in a file and keep the model swappable.

Frequently Asked Questions

Is the DALL-E AI image generator free?

DALL-E itself isn’t available anymore, free or paid. The closest free option is Bing Image Creator, which now runs Microsoft’s MAI-Image models with 15 fast creations a day. ChatGPT Images 2.5 is available to all ChatGPT users, according to OpenAI.

Can I still use DALL-E 3?

Not from OpenAI. It left ChatGPT in March 2025 and was removed from the API on 12 May 2026. API calls naming dall-e-3 no longer work. Use gpt-image-2.5-flare instead.

Is dall-efree.com made by OpenAI?

No. Its own FAQ calls it “an independent AI image generation platform”, and it offers models from many vendors. Treat any non-openai.com site with DALL-E in its name as a third-party wrapper.

What is the best alternative to the DALL-E AI image generator?

For conversational one-offs, ChatGPT Images 2.5. For free experiments, Bing Image Creator. For automation, an API model you can pin and swap: GPT-Image-2.5 Flare if you want OpenAI, or flux-2-pro on Replicate, which is what this site runs. If you want a wider tour, my AI image generator guide covers the broader field.

Do OpenAI images carry a watermark?

OpenAI says images made with its tools include C2PA metadata and invisible watermarking. That’s provenance data, not a visible logo on the image.

Will my old DALL-E prompts work on the new models?

Mostly, yes. Subject, style, light and composition still drive results. Strip out DALL-E-specific size tags (like 1792×1024) and parameters such as style: vivid, and keep “no text” in prompts where you don’t want lettering.

Final Thoughts

The DALL-E AI image generator had a great run. It made text-to-image normal, and plenty of people still say “DALL-E” the way people say “Google it”. But in October 2026 the model is retired, the free Bing route runs on other models, and most of the “DALL-E” sites ranking in Google sell access to everything except DALL-E.

Here’s what I’d do. For a handful of images, open ChatGPT or Bing and get on with your day. If images need to appear in your business without you, skip the hunt for the perfect model. Build the pipeline: one swappable model line, compression, a CDN, format rules, a log, and a loud failure when any step breaks. Mine has made 626 images this year and none of them needed me awake. When the next model gets retired, and it will, I’ll change one line.

If you’d rather have that built for you than build it yourself, that’s exactly the kind of done-for-you automation work I take on. In the meantime, the playbook below has the notes.