MMOOZZGG.com

Glossary

What is DALL-E?

In short

DALL-E is a family of OpenAI models that created images from a text description. In May 2026 both latest versions were turned off in the API.

Models and products · DALL-E

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine an artist who draws anything you say: “an armchair shaped like an avocado” or “a penguin in a hat”. It was one of the first such artists for computers, and newer, more skilled ones came after it. The old artist retired, but many learned from its drawings. This often happens with programs: new ones replace them, and you need to be ready for it.

In plain words

It was OpenAI’s line of picture generators. You described a scene in words, and the model drew the image in any style, even with objects and combinations that do not exist. The first version was described in a research paper in February 2021, and the second in April 2022. On 14 November 2025 OpenAI warned developers that the models would be retired, and on 12 May 2026 dall-e-2 and dall-e-3 were turned off in the API. As a replacement, the company offered the GPT Image family of models. For example, a greeting card service used to get illustrations from the model based on a holiday description; now another generator does the same job.

For business

While the models worked, companies connected them through the API for illustrations, banners and design options. Their retirement shows why you should never hard-wire an image generator into a process. Six months passed between the notice and the shutdown, and two of the three suggested replacements, gpt-image-1 and gpt-image-1-mini, also leave the API in October and December 2026. OpenAI services were never officially offered in Russia: the country is not on the supported list, and access from unsupported places can lead to a blocked account. The company did not publish open weights for the second and third versions, so you cannot install them on your own server. Plan the change of provider before you adopt a model, not after the shutdown email.

For a specialist

The first model is described in the paper by Ramesh et al., “Zero-Shot Text-to-Image Generation” (2021). A transformer models text and image tokens as one sequence and generates the picture autoregressively. The second version (Ramesh et al., “Hierarchical Text-Conditional Image Generation with CLIP Latents”, 2022) works in two stages. A prior turns the caption into a CLIP image embedding, and a diffusion decoder builds the picture from it. This approach gives more variety and allows image variations and text-guided edits without fine-tuning. The authors compared an autoregressive and a diffusion prior and noted that the diffusion prior is more compute-efficient and gives better samples. In the API, the third version took text and returned an image; it did not support edits through the edits method. The replacement for both models is listed in OpenAI’s table of deprecated models.

Example

A greeting card service generated pictures with the third version and, after the shutdown, rewrote its integration for another API. Now the model name lives in the settings, not in the code.

Related terms

Questions and answers

Does DALL-E still work? No, the dall-e-2 and dall-e-3 models were turned off in the API on 12 May 2026. OpenAI offered gpt-image-2, gpt-image-1 or gpt-image-1-mini as replacements, but the last two also leave the API before the end of 2026.
How did the second version differ from the first? The first generated an image as a sequence of tokens with a transformer. The second first predicted a CLIP image embedding from the text, and then a diffusion model built the picture from it.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us