AI Wiki

DALL·E Custom GPTs

ChatGPT
19 min read
Updated
Suggest editHistoryTalk
RawGraph

Last edited

Fact-checked

In review queue

Sources

29 citations

Revision

v3 · 3,864 words

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

See also: Custom GPTs, GPT Store, ChatGPT, DALL-E, and GPT Image 1

DALL·E Custom GPTs are Custom GPTs in the DALL·E category of the OpenAI GPT Store, the browse category the store launched with on January 10, 2024 for GPTs that exist primarily to generate or edit images [1][2]. At launch and throughout 2024 these GPTs called DALL-E 3 under the hood [1][3]. After OpenAI shipped native image generation in GPT-4o on March 25, 2025, image GPTs in ChatGPT switched to the new gpt-image-1 model, and that line has since advanced to gpt-image-1.5 (December 2025) and gpt-image-2 (April 2026), the model that powers ChatGPT image generation as of mid-2026 [6][23][25]. The DALL·E browse category kept its name even after the underlying model changed.

The official DALL·E GPT built by OpenAI, addressed in the store as @dall-e, remained one of the highest-traffic GPTs in the store throughout 2024 and into 2025 and acted as the canonical reference implementation for image-only GPTs. Custom GPTs and the GPT Store still exist in 2026, but OpenAI's platform focus has shifted toward apps in ChatGPT and agent-style successors introduced at DevDay in October 2025 [22][24].

Overview

A Custom GPT in the DALL·E category is a saved version of ChatGPT with a tailored system prompt, an optional knowledge file, and access to the image generation tool. Builders use the GPT Builder interface to set a persona (logo designer, coloring-book illustrator, meme maker, pixel artist, photo restorer), an output style, and constraints such as aspect ratio defaults. The GPT then calls the model that backs image generation in ChatGPT at that moment, which was DALL-E 3 from October 2023 to late March 2025, then gpt-image-1 and its successors: gpt-image-1.5 from December 2025 and gpt-image-2 from April 2026 [6][25].

The category is meaningful for three reasons. First, image generation is the most common use case for Custom GPTs after writing and research, and the DALL·E category is where most of that traffic lives. Second, the category is a useful proxy for OpenAI's image model rollouts: when the underlying model changes, every GPT in the category changes with it. Third, the category was one of the subject categories that defined the store's structure at launch, alongside Writing, Productivity, Research and Analysis, Education, Lifestyle, and Programming, sitting beneath a Featured row and a Trending leaderboard [1].

What image model powers ChatGPT image generation now?

As of mid-2026, image generation in ChatGPT, and therefore every GPT in the DALL·E category, is powered by gpt-image-2, which OpenAI markets as "ChatGPT Images 2.0" [25][26]. OpenAI announced it on April 21, 2026 and rolled it out to all ChatGPT and Codex users the next day [26]. gpt-image-2 is the first OpenAI image model with reasoning, or "thinking," built into generation: it can search the web, produce several candidate images from a single prompt, and check its own output before returning it [26]. It renders text far more reliably than earlier models, including non-Latin scripts such as Japanese, Korean, Hindi, and Bengali, and can output up to 2K resolution [26]. OpenAI described the release as bringing "unprecedented specificity and fidelity, able to follow instructions and render fine-grained elements that often break models" [25].

gpt-image-2 is the fourth model in the line that replaced DALL-E 3. The sequence runs gpt-image-1 in March 2025, a cheaper gpt-image-1-mini at DevDay in October 2025, gpt-image-1.5 in December 2025, and gpt-image-2 in April 2026 [6][23][25]. DALL-E 3 itself remained available in the API for a period before OpenAI moved to deprecate it [20], which is part of why the store category still carries the DALL·E name.

DALL·E lineage and integration into ChatGPT

The lineage of image models that have powered the DALL·E category of the GPT Store goes back to January 2021, well before the store existed [4][5].

ModelReleaseArchitectureNotes
DALL-EJanuary 5, 202112B parameter autoregressive transformer derived from GPT-3Research preview. Generated 256x256 images from text. Not publicly accessible at first.
DALL-E 2April 6, 2022 (announced); September 28, 2022 (public)3.5B parameter diffusion model conditioned on CLIP image embeddingsBeta opened to 1 million waitlisted users on July 20, 2022. Waitlist removed September 28, 2022. API followed in early November 2022.
DALL-E 3September 20, 2023 (announced); October 19, 2023 (in ChatGPT)Diffusion model trained on highly descriptive synthetic captionsIntegrated natively into ChatGPT Plus and ChatGPT Enterprise. API and Labs availability followed in November 2023. Powered every GPT in the DALL·E category at launch.
gpt-image-1March 25, 2025 (in ChatGPT); April 23, 2025 (API)Autoregressive multimodal model integrated into GPT-4oReplaced DALL-E 3 as the native image model in ChatGPT. Marketed in the ChatGPT product as "4o Image Generation" or "ChatGPT Images". Native image inputs enabled true image-to-image editing for the first time [7][21].
gpt-image-1-miniOctober 6, 2025 (DevDay 2025)Smaller, cheaper variant of gpt-image-1Roughly 80% cheaper than gpt-image-1 through the API. Aimed at high-volume and cost-sensitive workloads [6].
gpt-image-1.5December 16, 2025Successor to gpt-image-1 (autoregressive)More precise edits with better preservation of logos and faces, better instruction following, and improved rendering of dense and small text. About 20% cheaper than gpt-image-1 for API image inputs and outputs. Rolled out to all ChatGPT users, including free accounts, during OpenAI's "Little Shipmas" [23].
gpt-image-2April 21, 2026 (announced); April 22, 2026 (all ChatGPT and Codex users)Autoregressive multimodal model with built-in reasoningMarketed as "ChatGPT Images 2.0". Adds "thinking" (web search, multiple candidate images, self-verification), much stronger text rendering including non-Latin scripts, and output up to 2K resolution. Current default image model in ChatGPT [25][26].

DALL-E 3 reached ChatGPT in two stages. OpenAI announced the model on September 20, 2023, and rolled it out to ChatGPT Plus and Enterprise subscribers on October 19, 2023 [3]. API access followed in early November. This is the model that the GPT Store was built on. When the store opened on January 10, 2024, every image GPT in the catalogue was a thin wrapper around DALL-E 3, with the builder's system prompt nudging the model toward a specific style or workflow [1].

The shift to gpt-image-1 in March 2025 changed the picture without changing the category name. OpenAI revealed the new model as "GPT-4o image generation" in a blog post on March 25, 2025, made it the default image generator in ChatGPT a few days later, and opened it on the API on April 23, 2025 under the identifier gpt-image-1 [7][21]. According to OpenAI, more than 130 million users created over 700 million images in the first week, and Sam Altman said on social media that the company's GPUs "were melting" from the level of use [7][11]. The model has continued to evolve since: gpt-image-1.5 in December 2025 and gpt-image-2 in April 2026 both shipped straight into the same generic image tool that DALL·E-category GPTs call, so every GPT in the category inherited the upgrades automatically [23][25].

GPT Store DALL·E category

The GPT Store launched on January 10, 2024 with several built-in subject categories visible in the browse interface, plus a Featured row and a Trending leaderboard [1][2]. DALL·E was one of those categories. The others OpenAI highlighted were Writing, Research and Analysis, Programming, Education, and Lifestyle, with a Productivity category also surfaced in the store UI [1]. At launch the store featured GPTs from official partners, including AllTrails (a hiking-trail recommender), Khan Academy (a code tutor called Code Tutor), and Canva (a design assistant called Canva GPT). The DALL·E category leaned heavily on community-built GPTs. OpenAI said that in the roughly two months after GPTs were announced in November 2023, users had already created more than three million custom versions of ChatGPT [1].

Access to the store has always been gated by a paid ChatGPT plan. At launch that meant Plus, Team, or Enterprise [1][2]. ChatGPT Pro, announced in December 2024, was added to the list. Free users could not use Custom GPTs at all until limited free-tier access arrived later in 2024, and even then image generation in GPTs remained subject to usage caps.

In March 2024 OpenAI launched the GPT builder revenue program, paying U.S. builders based on user engagement with their GPTs. Image GPTs were eligible. The exact payout formula was not made public, but builders reported that engagement, not raw chat count, drove the figure.

Do Custom GPTs and the GPT Store still exist in 2026?

Yes. Custom GPTs and the GPT Store are still available in 2026, and the DALL·E category still appears in the store, but OpenAI's platform strategy has moved on. At DevDay on October 6, 2025, OpenAI introduced apps in ChatGPT and an Apps SDK built on the Model Context Protocol (MCP), a more capable successor to GPTs that renders interactive interfaces inside a chat and can trigger real actions such as ordering groceries or turning an outline into a slide deck [22][24]. Launch partners included Spotify, Canva, Zillow, Figma, Coursera, Expedia, and Booking.com [22]. On December 18, 2025, OpenAI opened app submissions to all developers, effectively launching a second, app-based store alongside the original GPT Store [24].

For individual ChatGPT users on Free, Plus, and Pro plans, existing Custom GPTs, including DALL·E-category image GPTs, continue to work. OpenAI has been steering business and enterprise customers toward newer agent-style tools, and the GPT Store, while still live and hosting the hundreds of thousands of public GPTs built since 2024, is widely described as having entered maintenance mode as new investment flows into apps and agents [24]. For a builder whose goal is simply "generate images in a conversation," a DALL·E-category Custom GPT still does the job, and it now does it with gpt-image-2 rather than DALL-E 3.

Notable DALL·E-category GPTs

The table below lists DALL·E-category GPTs that are verifiable on the live OpenAI GPT Store catalogue at chatgpt.com/g/. Popularity rankings shift constantly because the store sorts on a recency-weighted chat count, so any single ranking is a snapshot. The descriptions reflect the GPT's stated purpose, not a quality judgment.

GPTBuilderPurpose
DALL·EOpenAIThe official, baseline image GPT. Used as the default image generator inside ChatGPT before March 2025.
Image GeneratorVarious community builders use this nameGeneral text-to-image with sensible defaults for prompt formatting.
Logo CreatorCommunityGenerates logo concepts. Tries to keep text spelled correctly, a known weakness of DALL-E 3.
Coloring Book HeroCommunityOutputs black-and-white line art suitable for printable coloring pages, popular with Amazon KDP self-publishers.
Hot ModsOpenAITakes an uploaded photo and modifies it (cartoonizes, swaps backgrounds, restyles characters). One of the few OpenAI-built GPTs in the category.
Cartoonize YourselfCommunityPhoto-to-cartoon stylization, a precursor to the Ghibli-style trend that exploded after gpt-image-1 launched.
Pixel Art GeneratorCommunityGenerates retro pixel-art sprites and tilesets. Has had to adapt prompts heavily because DALL-E 3 tended to anti-alias edges.
Image Edit and img2imgCommunityImage-to-image workflows. Much more effective after the March 2025 switch to gpt-image-1, which natively accepts image inputs.
GradientCommunityImage and art generator, popular for stylized illustration work.

A few cautions about this list. The store does not publish a stable canonical ID for many of these GPTs, and several have name collisions: there are at least three different "Logo Creator" GPTs that have been featured at different times. Third-party aggregator sites like gptstore.ai mirror the catalogue but rank by their own scraped metrics, not OpenAI's. The only authoritative source is the live store at chatgpt.com/gpts, which is why this list deliberately stays short.

Transition to gpt-image-1 and the Ghibli viral wave

The transition from DALL-E 3 to gpt-image-1 inside the GPT Store happened in stages and was not announced loudly. OpenAI unveiled native GPT-4o image generation in a blog post on March 25, 2025, branded internally as "ChatGPT Images" [7]. Within days, the default image generator inside ChatGPT switched from DALL-E 3 to the new model. The official DALL·E GPT continued to exist as a separate entry point, but every other image GPT in the store quietly inherited the new model, because they all call the same generic image tool exposed to GPTs.

The most visible side effect was the Studio Ghibli style wave that started around March 26, 2025 [8]. Users discovered that the new model was much better than DALL-E 3 at replicating the soft, painterly look of films from Studio Ghibli, and within a day social media was full of Ghibli-style portraits of family photos, news events, and political figures [9]. Sam Altman changed his profile picture on X to a Ghibli-style avatar [11]. The White House posted a Ghibli-style image about an immigration enforcement action, which prompted a public response from film distributor GKids that compared the AI filter to its concurrent IMAX re-release of Princess Mononoke [10].

The trend revived an old quote from Studio Ghibli co-founder Hayao Miyazaki, who said in a 2016 NHK documentary that AI-generated animation was "an insult to life itself" [8][9]. OpenAI did not block the style outright, although it did refuse prompts that asked for the work of named living artists, a restriction that had existed since DALL-E 3 launched [10].

For the DALL·E category of the GPT Store, the practical effect of the switch was that image-to-image workflows became dramatically more usable. DALL-E 3 could not really edit an uploaded image; it had to regenerate from text. gpt-image-1 accepts an image as input and edits it, which means GPTs like Hot Mods, Cartoonize Yourself, and Image Edit and img2img suddenly worked the way users had always expected them to [21]. Later models pushed this further: gpt-image-1.5 was tuned specifically for precise edits that preserve logos and faces, and gpt-image-2 added the ability to reason about an edit before committing to it [23][26].

Restrictions and policies

Image generation in ChatGPT, and therefore every GPT in the DALL·E category, is bound by OpenAI's content policy. The main restrictions, as documented in OpenAI's safety pages and reported by TechCrunch, the Washington Post, and Wired, include:

  • Named public figures and realistic likenesses. The model refuses to generate images of named real people in most cases. The refusal expanded over time. In 2023 it covered prominent political figures; by 2024 it covered most celebrities; after the Ghibli wave OpenAI tightened the rules again in April 2025 to handle uploaded portraits more carefully [10].
  • Trademarked characters. Mickey Mouse, Marvel characters, Pokemon, and similar named IP are blocked or aggressively rewritten. Enforcement is inconsistent. Users have reported that lookalikes sometimes slip through, while exact names almost always trigger a refusal [8].
  • Living artists by name. Asking for an image "in the style of [living artist]" is refused, with the exception of widely used stylistic terms like "anime" or "watercolor". The Ghibli loophole worked because "Studio Ghibli" is a studio name, not a living individual; OpenAI did not change that behavior even at the peak of the trend [10].
  • Adult content, graphic violence, and hate. Standard categories. Refusals are stricter inside Custom GPTs than in the base ChatGPT for some categories, because GPTs are public artifacts subject to a separate moderation review.
  • C2PA Content Credentials. Starting February 2024, all images generated by DALL-E 3 in ChatGPT and the API carry C2PA metadata that records the model used and that an OpenAI service produced the file [12][13]. The metadata survives normal file copying but is stripped by screenshots and by some image-editing pipelines. gpt-image-1 and its successors inherited the same scheme [21]. ChatGPT-generated images also include an additional manifest indicating ChatGPT origin.

C2PA is an open standard cofounded by Adobe, Microsoft, Intel, BBC, Truepic, and Arm. OpenAI joined the C2PA steering committee in 2024 [14]. The watermark is two parts: an optional visible "CR" credentials icon in the image corner, and the invisible signed metadata embedded in the file.

How do DALL·E GPTs compare to other image-generation tools?

The DALL·E category competes with a handful of standalone tools that sit outside ChatGPT. The table below summarizes how they stack up for the kinds of work that DALL·E-category GPTs are typically used for [16].

ToolOwnerStrengthsWeaknesses relative to DALL·E in ChatGPT
MidjourneyMidjourney, Inc.Best photorealism and cinematic aesthetic since v6. Strong character reference (cref). Discord and web interface.No real conversational interface. Less reliable on multi-subject prompts. Discontinued free tier in 2024; paid plans start at $10/month.
Stable DiffusionStability AI and the open-weights communityOpen weights. Runs locally. Fine-tuning, LoRA adapters, ControlNet, pose and depth control. Free.No conversational interface unless wrapped. Requires GPU hardware or hosted services. Higher learning curve.
FluxBlack Forest LabsBuilt by former Stability AI researchers who created the original Stable Diffusion. Strong text rendering and prompt adherence. FLUX.1 [pro] (API), FLUX.1 [dev] (open weights non-commercial), FLUX.1 [schnell] (Apache 2.0).No first-party chat interface. Accessed via fal.ai, Replicate, or other third parties [17][18].
Adobe FireflyAdobeIndemnification: Adobe says it will defend commercial users against claims that outputs infringe copyright. Trained on Adobe Stock and licensed content. Integrated into Photoshop and Illustrator.Less raw aesthetic quality than Midjourney or gpt-image-1. Tied to Adobe Creative Cloud subscriptions.
IdeogramIdeogram AIBest in class for rendering legible text inside images: logos, posters, banners.Narrower stylistic range than competitors.
RecraftRecraftVector output, background removal, mockups, and reusable brand workflows.Smaller community and fewer GPT-style wrappers.

DALL·E-category GPTs win on ease of use and on the conversational workflow. You can describe what you want in plain English, see candidate images, and ask the GPT to refine specific parts, all inside the same ChatGPT thread that holds the rest of your project context. Midjourney and Stable Diffusion produce better images for many tasks, but neither has anything like the same conversational loop. After the move to gpt-image-1 in March 2025, and further with gpt-image-1.5 and gpt-image-2, the gap on raw image quality narrowed enough that the conversational advantage often matters more than the pixel-level comparison.

One place where DALL·E GPTs historically lost is text inside images. DALL-E 3 was notoriously bad at spelling, and even gpt-image-1 lagged Ideogram and Flux on long stretches of in-image text. That gap has narrowed sharply: OpenAI made improved rendering of dense, small, and non-Latin text a headline feature of both gpt-image-1.5 and gpt-image-2, the latter of which uses its reasoning step to double-check spelling before returning an image [23][26]. Logo Creator GPTs that once worked around the problem by generating logos with no text and telling the user to add type in Canva or Figma afterward increasingly no longer need to.

Beyond images: Sora and the wider OpenAI media stack

The DALL·E category sits in a wider OpenAI media stack that grew quickly in 2024 and 2025. Sora, OpenAI's text-to-video model, launched publicly as Sora Turbo on December 9, 2024 at sora.com, with access included in ChatGPT Plus and ChatGPT Pro plans in the U.S. and Canada [15][19]. The EU and U.K. were excluded at launch. OpenAI released Sora 2, its flagship video-and-audio model with synchronized dialogue and sound effects, on September 30, 2025, alongside a standalone Sora iOS app that opened by invite in the U.S. and Canada [27].

The Sora consumer product proved short-lived. OpenAI announced a two-stage shutdown in 2026: the Sora app and web experience were discontinued on April 26, 2026, and the Sora API is scheduled to be discontinued on September 24, 2026, ending Sora as a consumer product [28][29]. OpenAI framed the move as a strategic pivot that funnels compute toward coding tools and enterprise customers, and said Sora would continue as a research effort on world models aimed at "automating the physical economy" [29]. Sora was a separate product, not a GPT, but for a time it was the obvious next step for many DALL·E-category builders, and the same Content Credentials and policy framework applied to it. Watermarks on Sora videos were visible in the bottom-right corner by default.

Even with the Sora consumer app winding down, builders in the DALL·E category increasingly think in multimodal terms: a single project might use DALL·E-category GPTs for stills and a separate video model for short clips, with ChatGPT coordinating both.

See also

References

  1. OpenAI, "Introducing the GPT Store", January 10, 2024. ↩
  2. TechCrunch, "OpenAI launches a store for custom AI-powered chatbots", January 10, 2024. ↩
  3. OpenAI blog, "DALL·E 3 is now available in ChatGPT Plus and Enterprise", October 2023. ↩
  4. OpenAI blog, "DALL·E: Creating images from text", January 5, 2021. ↩
  5. Wikipedia, "DALL-E", retrieved 2026. ↩
  6. Wikipedia, "GPT Image", retrieved 2026. ↩
  7. OpenAI blog, "Introducing 4o Image Generation" / "The new ChatGPT Images is here", March 25, 2025. ↩
  8. TechCrunch, "OpenAI's viral Studio Ghibli moment highlights AI copyright concerns", March 26, 2025. ↩
  9. CNN, "Viral Studio Ghibli-style AI images showcase power and copyright concerns of update", March 27, 2025. ↩
  10. Washington Post, "AI generated Ghibli images go viral as OpenAI loosens its rules", March 28, 2025. ↩
  11. Variety, "OpenAI CEO Responds to ChatGPT Users Creating Studio Ghibli-Style AI Images", March 2025. ↩
  12. OpenAI Help Center, "C2PA in ChatGPT Images". ↩
  13. PetaPixel, "AI Images Generated on DALL-E Now Contain the Content Authenticity Tag", February 8, 2024. ↩
  14. C2PA, "OpenAI Joins C2PA Steering Committee", 2024. ↩
  15. TechCrunch, "OpenAI's Sora video generator is launching for ChatGPT Pro and Plus subscribers", December 9, 2024. ↩
  16. Creative Bloq, "Midjourney vs Dall-E 3 vs Stable Diffusion: which AI image generator is best?". ↩
  17. VentureBeat, "Stable Diffusion creators launch Black Forest Labs, secure $31M for FLUX.1 AI image generator", 2024. ↩
  18. Wikipedia, "Flux (text-to-image model)", retrieved 2026. ↩
  19. Wikipedia, "Sora (text-to-video model)", retrieved 2026. ↩
  20. OpenAI Developer Community, "OpenAI is making a huge mistake by deprecating DALL-E-3", 2025. ↩
  21. OpenAI, "Introducing our latest image generation model in the API" (gpt-image-1), April 23, 2025. ↩
  22. OpenAI, "Introducing apps in ChatGPT and the new Apps SDK", October 6, 2025. ↩
  23. OpenAI Developer Community, "GPT-Image-1.5 rolling out in the API and ChatGPT", December 16, 2025. ↩
  24. TechCrunch, "ChatGPT launches an app store, lets developers know it's open for business", December 18, 2025. ↩
  25. OpenAI, "Introducing ChatGPT Images 2.0", April 21, 2026. ↩
  26. TechCrunch, "ChatGPT's new Images 2.0 model is surprisingly good at generating text", April 21, 2026. ↩
  27. OpenAI, "Sora 2 is here", September 30, 2025. ↩
  28. OpenAI Help Center, "What to know about the Sora discontinuation", 2026. ↩
  29. The Decoder, "OpenAI sets two-stage Sora shutdown with app closing April 2026 and API following in September", 2026. ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

2 revisions by 1 contributors · full history

Suggest edit