DALL·E 2

What DALL·E 2 was

DALL·E 2 was OpenAI’s second-generation image-generation model. It transformed natural-language descriptions into original images and artwork, allowing a user to describe subjects, visual attributes, compositions, and artistic styles in a prompt.

OpenAI introduced DALL·E 2 as a research preview in April 2022. The provider described it as producing more realistic and accurate images than the original DALL·E, with four times greater resolution. That comparison is an OpenAI claim about the model’s improvement over its predecessor, rather than an independent performance assessment.

The model’s API identifier was dall-e-2. It was historically important because it extended image generation beyond simple text-to-image requests into workflows that could modify an existing image or produce related alternatives.

Core image capabilities

DALL·E 2 supported text-to-image generation: a user could provide a written description and request a new raster image. For example, a prompt could describe a particular object, setting, visual style, or combination of concepts. The model was designed to combine concepts, visual attributes, and artistic styles in one generated result.

It also supported image-based workflows:

• Editing: modifying an existing image according to a prompt.
• Inpainting: replacing or filling a selected area within an image.
• Outpainting: extending an image beyond its original boundaries.
• Variations: generating alternative images based on an existing image.

These capabilities made DALL·E 2 useful for visual exploration. A designer could test multiple compositions, extend a scene, replace part of an image, or use a source image as the basis for variations rather than starting from a blank canvas every time.

Inputs, outputs and supported formats

DALL·E 2 accepted text prompts for image generation and image inputs for editing or variation workflows. Its output was a raster image. It did not produce text, audio, video, embeddings, structured data, tool calls, or general-purpose language-model responses.

Documented generation sizes were 256×256, 512×512, and 1024×1024 pixels. The documented prompt limit for the image-generation API was 1,000 characters. The supplied specifications do not identify a context window, maximum output-token limit, knowledge cutoff, streaming support, function calling, or batch API capability; those language-model concepts generally do not apply to this image-only model.

Historical API pricing

DALL·E 2 was priced per generated image rather than through a recurring subscription tier. Historical API prices were:

• $0.016 per 256×256 image
• $0.018 per 512×512 image
• $0.020 per 1024×1024 image

These figures are historical and should not be treated as current purchase prices. The model is no longer available for new API requests, so there is no current production price or billing plan for DALL·E 2.

The OpenAI API public beta for DALL·E was announced on November 3, 2022. At that time, per-image pricing provided a straightforward cost model: larger output dimensions cost slightly more, while smaller images were cheaper when high resolution was not required.

Retirement and position in OpenAI’s lineup

OpenAI deprecated the dall-e-2 model and removed it from the API on May 12, 2026. The retirement means that DALL·E 2 is now relevant primarily for historical research, legacy documentation, and comparisons with newer image-generation systems.

OpenAI recommended moving to newer GPT Image models, including GPT Image 2, GPT Image 1, or GPT Image 1 Mini. The available research identifies those models as successors for new image-generation work, but it does not provide a detailed feature or pricing comparison for each one. Readers evaluating a current OpenAI image model should therefore treat GPT Image 2 as a newer alternative rather than attempting to build a new application around DALL·E 2.

Strengths and limitations

DALL·E 2’s main strength was the breadth of its image workflow for its generation. It did more than create a new image from a caption: it supported editing, inpainting, outpainting, and variations. That made it useful for iterative visual work, such as changing a selected object, extending a background, or exploring several interpretations of a source image.

Its other historical advantage was a simple API pricing structure with three documented output sizes. Users could choose a lower-cost 256×256 or 512×512 image when a smaller asset was sufficient, or request 1024×1024 output for a larger result.

Its limitations are significant for current use. DALL·E 2 was an earlier-generation image model and lacked the newer quality levels, resolutions, capabilities, and workflow integrations associated with later GPT Image models. It was not a reasoning or coding model, did not support general conversational responses, and did not provide audio or video generation.

Most importantly, retirement is now a practical limitation: even if its historical capabilities fit a project, the API cannot be used for a new production integration after the shutdown date.

Reasoning, coding and tool support

DALL·E 2 did not provide general-purpose reasoning, coding, or tool-use capabilities. A text prompt could describe the image to create, but the model did not return an explanation, execute code, call external functions, browse the web, or produce structured JSON.

This distinction matters when choosing a model. DALL·E 2 could be part of a larger application in which another system handled conversation, prompt construction, workflow logic, or post-processing, but DALL·E 2 itself was an image-generation component rather than an all-purpose assistant.

When to choose this model

For new deployments, there is no reason to choose DALL·E 2 because OpenAI has removed it from the API. A current GPT Image model is the appropriate direction for a new OpenAI image-generation integration.

DALL·E 2 can still be the right subject for a historical comparison or a legacy migration project. It is especially relevant when documenting:

• the development of OpenAI’s image-generation products;
• early text-to-image API pricing and output-size choices;
• the introduction of inpainting, outpainting, and image variations in an OpenAI workflow;
• the behavior of an existing application that was built around the dall-e-2 model identifier.

If the requirement is current image quality, continued API access, newer editing workflows, or a supported production service, a newer GPT Image model is more appropriate. If the requirement is text reasoning, coding, structured output, or tool use, an image model is the wrong category of option and a current language model should handle those tasks separately.

Bottom line

DALL·E 2 was a milestone image-generation model that combined text-to-image creation with editing, inpainting, outpainting, and variations. Its documented API supported three square image sizes and historically charged between $0.016 and $0.020 per image. Those facts are useful for understanding OpenAI’s earlier image API, but the model is no longer a current service: it was deprecated and removed from the API on May 12, 2026. New projects should use a supported GPT Image model instead.