DALL-E 3 Is Retired: What Developers Should Use Now

DALL-E 3 is over, but this page still matters

DALL-E 3 is no longer an available OpenAI API model. OpenAI removed both the dall-e-2 and dall-e-3 model snapshots from the API on May 12, 2026. Its current documentation recommends GPT Image 2 for new image generation and editing, with GPT Image 1 and GPT Image 1 Mini also named as replacements in the original retirement notice.

That sentence should be the first thing anyone learns about DALL-E 3 in 2026. A directory page that still advertises active API access, old per-image prices, or ChatGPT subscriptions as DALL-E 3 plans is not merely behind the times. It can send a developer toward a model that no longer answers requests and make a buyer compare prices for a product that cannot be purchased.

Retirement does not make the record useless. DALL-E 3 influenced how people expected text-to-image systems to behave. It was known for translating long descriptions into coherent scenes, and its integration with conversational prompting made visual generation feel less like operating a specialist model. Thousands of prompts, internal tools, tutorials, screenshots, and product requirements still refer to it. The useful page now is one that helps readers understand that history and move forward without pretending the old doorway remains open.

A clean ending is more useful than fake availability

AI directories face an awkward choice when a famous model disappears. Deleting the page removes a useful answer for people searching old documentation, debugging an integration, or asking why familiar code stopped working. Leaving the page untouched is worse because it creates confident misinformation. The responsible middle path is to keep the identity, mark the status plainly, and rebuild the content around migration.

That is the purpose of this DALL-E 3 record. It remains indexed because the name still carries meaning, but its decision box should not position it beside live products as though access were equal. Its old prices belong in history, not in a current buying table. Its workflow should begin with inventory and replacement planning, not instructions to sign up and generate an image.

The date matters. OpenAI announced the deprecation to affected developers in November 2025 and set May 12, 2026 as the shutdown date. As of this update, that date has passed. This is not a warning about a future change or an inference from a missing button. OpenAI's current model page labels DALL-E 3 as deprecated and removed from the API.

There is a second timing detail worth noticing. OpenAI now presents GPT Image 2 as its current image model, and it has also scheduled GPT Image 1 Mini, GPT Image 1.5, and the chatgpt-image-latest alias for removal on December 1, 2026. A migration that moves from DALL-E 3 to another model already approaching retirement only postpones the same work. New production systems should evaluate GPT Image 2 first unless a verified constraint requires another target.

What DALL-E 3 actually did

DALL-E 3 accepted text and returned an image. Its API generation contract allowed prompts up to 4,000 characters, one generated image per request, and three output dimensions: 1024 by 1024, 1792 by 1024, and 1024 by 1792. Users could choose standard or HD quality and select either a vivid or natural style. Responses could contain a temporary URL or base64-encoded image data, and the API exposed the revised prompt that the model used.

That revised-prompt behaviour was central to the experience. A user could provide a natural description while the system expanded or adjusted the wording before generation. For casual creation, this often made the model feel more cooperative. For a production system, it also meant the submitted prompt was not always the complete explanation of the output. Anyone trying to recreate an archived DALL-E 3 result should preserve both the original prompt and the returned revised prompt when available.

DALL-E 3 was a generation model, not a modern image-editing workflow. Old pages sometimes blur that distinction because OpenAI's Images API also contained edit and variation capabilities associated with other models. The current GPT Image family accepts image inputs and supports generation as well as edits. That is a materially different contract, not a renamed version of the same endpoint behaviour.

The old model also had a simple price story. API charges were quoted per generated image, varying by size and quality. Those historical numbers can help explain old invoices, but they no longer describe an available purchase. Current GPT Image pricing is token based, which changes how teams estimate, instrument, and optimise image jobs.

The migration is a product rewrite, not a model swap

Changing the model string is the smallest part of a safe migration. An application can compile after that edit and still be wrong in almost every operational detail. GPT Image models add image inputs, edit operations, different quality names, different size options, new output controls, token usage, and different response handling. The current Image API also returns GPT Image outputs as base64 data rather than following the old temporary URL pattern.

Start with the request schema. DALL-E 3 used standard or HD quality, while GPT Image uses low, medium, high, or automatic quality where supported. DALL-E 3 offered vivid and natural styles; that parameter is not the general control surface for GPT Image 2. DALL-E 3 had three fixed sizes. GPT Image 2 supports standard dimensions and a much more flexible resolution system, subject to divisibility, aspect-ratio, pixel, and edge limits described in the current API reference.

Then inspect the response path. Code that expects a URL, downloads it within its validity window, and stores the result will not necessarily work with a base64 image payload. The new response may also carry usage information that should be logged for cost analysis. File extension, MIME type, transparency, compression, and output format now deserve explicit handling instead of assumptions inherited from the older model.

Finally, examine the surrounding product. Retry logic, timeouts, rate limits, moderation handling, image storage, cache keys, analytics, and user-facing progress states may all depend on old behaviour. If a system exposes vivid and natural as customer settings, the team must decide whether to remove those controls, translate them into prompt guidance, or replace them with a more relevant choice. Silently ignoring an obsolete option is rarely a good user experience.

GPT Image 2 changes the creative contract

OpenAI describes GPT Image 2 as its current state-of-the-art image generation model. It accepts text as input and images as both input and output, supports generation and editing, and works through the Image API. Image generation can also be used through the Responses API as a built-in tool for conversational and multi-step experiences.

The addition of image inputs is not a minor feature. A product can now preserve a reference composition, revise a product photograph, change a background, iterate on an approved asset, or carry visual context through a conversation. DALL-E 3 workflows often treated each generation as a new attempt described in words. GPT Image workflows can treat an existing image as part of the working state.

Flexible dimensions also reduce some awkward production steps. GPT Image 2 supports arbitrary resolutions within documented constraints, as well as familiar standard sizes. The API reference notes that width and height must be divisible by 16, the aspect ratio must remain between 1:3 and 3:1, and higher resolutions carry additional experimental and maximum limits. Applications should validate these rules before sending a request rather than passing arbitrary user values to the API.

Output control is broader too. GPT Image models support PNG, JPEG, and WebP, with compression controls for relevant formats and transparent backgrounds where supported. These options create better production paths, but each one adds a decision. A transparent PNG may be right for a product cutout and wasteful for a photographic background. JPEG or WebP may reduce delivery size, but compression should be checked against text, edges, and brand details.

Costs moved from per-image simplicity to token economics

OpenAI's current standard pricing for GPT Image 2 is expressed per one million tokens. Text input is listed at $5, cached text input at $1.25, image input at $8, cached image input at $2, and image output at $30. Batch pricing is lower, with the corresponding figures shown at $2.50, $0.625, $4, $1, and $15.

Those numbers are not a flat image price. The cost of a job depends on its text, image inputs, and output. Dimensions, quality, editing context, and workflow design can affect consumption. The official image guide therefore points developers to an image-generation calculator for estimates. A directory should not convert token rates into a single universal price without declaring assumptions.

For teams, the useful unit is approved output rather than generated output. A cheap image that needs ten reruns can cost more than a stronger first attempt. An edit that preserves an approved product may be more valuable than generating a complete scene again. Cost dashboards should connect token use to the job, model, quality, dimensions, result status, and eventual approval, not simply count API calls.

Batch pricing can be attractive when the work does not need an immediate response, such as preparing a catalogue overnight or generating a large evaluation set. Interactive editors, customer-facing previews, and conversational sessions usually need standard processing. The architecture should make that distinction deliberately rather than route every image through the same latency and price path.

Build a migration benchmark you can defend

A reliable migration begins with representative prompts, not a hand-picked gallery. Collect the jobs the old system actually performed: product scenes, illustrations, social graphics, backgrounds, concept art, diagrams, posters, or whatever mattered to the application. Include easy cases, frequent cases, valuable cases, safety-sensitive cases, and the prompts most likely to fail.

Preserve the DALL-E 3 evidence you still have. Store original prompts, revised prompts, old outputs, size and quality settings, application version, and any human approval notes. Do not claim deterministic recreation. The goal is to establish a meaningful baseline for the new system, not to force GPT Image 2 to imitate every historical pixel.

Define acceptance criteria before generating the comparison set. Prompt adherence, subject identity, product accuracy, readable text, layout, style, safety behaviour, latency, file size, transparency, and cost may all matter. Weight them according to the product. A marketing illustration and a packaging edit should not be scored with the same priorities.

Run the new model through a pinned request configuration and record every output. Human reviewers should see the prompt and relevant reference images but should not be steered by which system produced the result. Capture rejection reasons in structured categories. A vague note such as looks worse is much less useful than incorrect logo, missing object, distorted hand, unreadable text, unsafe composition, or colour mismatch.

The benchmark should also include integration behaviour. Test invalid sizes, unsupported parameters, moderation responses, network failures, timeouts, storage errors, and rate limits. Confirm that base64 decoding, format detection, metadata, and retention all work. Migration is complete only when the product handles both successful images and the messy edges around them.

Prompt rewriting and archived outputs

Teams that built prompt libraries around DALL-E 3 should resist the urge to copy every prompt unchanged and call the work finished. The old model's revised prompt could add detail or reinterpret instructions. GPT Image models have different capabilities and may respond differently to structure, references, editing directions, and constraints.

Separate the brief from the model adapter. The brief should express the durable intent: subject, purpose, audience, composition, required elements, forbidden changes, brand rules, output context, and acceptance criteria. A model-specific adapter can then translate that intent into parameters and prompting suitable for GPT Image 2. This makes the next migration less painful and prevents historical quirks from leaking into the product interface.

Archived DALL-E 3 images should be treated as records, not renewable API assets. Keep the original files in durable storage because old temporary URLs were never intended to last. Preserve provenance, prompt data, rights records, and approval history. If an image is used in a live campaign or product, retain the exact delivered master rather than relying on the possibility of regenerating it later.

It is also worth cleaning public documentation. Remove active DALL-E 3 signup instructions, obsolete price tables, old model-selection examples, and code that implies the model is callable. Add the retirement date and the chosen replacement. If an old tutorial remains valuable, label it as historical and link readers to the current image-generation guide.

The WhatAI verdict

DALL-E 3 deserves a place in the history of accessible image generation, but it no longer belongs in a current shortlist of API models. OpenAI has closed the endpoint. The honest recommendation for new development is GPT Image 2, with a migration that accounts for editing, image inputs, output formats, flexible dimensions, response handling, token pricing, and updated safety behaviour.

For a reader who simply wants to generate an image today, this record is a signpost toward current OpenAI image tools rather than a product to buy. For a developer maintaining old code, it is a checklist: find every dall-e-3 reference, preserve historical evidence, redesign the request and response path, benchmark representative jobs, measure cost per approved asset, and remove obsolete claims from documentation.

The broader lesson is useful beyond OpenAI. Model pages have a lifecycle. A responsible AI directory should be as clear about endings as it is enthusiastic about launches. DALL-E 3's value in 2026 is no longer the images it can generate. It is the migration knowledge that helps people leave it cleanly.

WhatAI Decision Box

OpenAI removed DALL-E 3 from the API on May 12, 2026. This updated record explains the shutdown, preserves the model's historical request and response behaviour, and gives developers a practical path to GPT Image 2.

The dall-e-3 model can no longer be used through the OpenAI API. GPT Image 2 is OpenAI's recommended current model for image generation and editing, with image inputs, flexible dimensions, broader output controls, and token-based pricing.

A safe migration must update quality and size parameters, response handling, storage, cost tracking, moderation, and evaluation. Teams should preserve old prompts and revised prompts, then benchmark representative production jobs against explicit acceptance criteria.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-08-17.

DALL-E 3 Community Discussions

Explore community discussions. Ask and answer questions on DALL-E 3 to grow and learn together.

DALL-E 3 Pros & Cons

How to Get Results with DALL-E 3: Step-by-Step Workflow

1. Confirm every DALL-E 3 dependency

   Search source code, environment variables, dashboards, queues, tests, documentation, pricing pages, user interfaces, and stored job metadata for dall-e-3 and its old parameters.

2. Preserve the historical evidence

   Archive original prompts, returned revised prompts, output files, size and quality settings, cost records, approval notes, and representative failures before changing the production path.

3. Select the supported target

   Evaluate GPT Image 2 first because OpenAI recommends it for current generation and editing. Confirm model access, organization verification, rate limits, regional requirements, and budget.

4. Map the request contract

   Replace DALL-E 3 quality, style, size, response-format, and one-image assumptions with validated GPT Image settings for quality, dimensions, formats, transparency, compression, and image inputs.

5. Rebuild response handling

   Decode and validate base64 output, detect the file format, store the image durably, capture token usage and metadata, and remove any dependency on temporary DALL-E image URLs.

6. Create a representative evaluation set

   Include frequent, valuable, difficult, safety-sensitive, and failure-prone jobs. Define prompt adherence, reference fidelity, text accuracy, composition, latency, safety, and cost criteria before running it.

7. Test operational failures

   Exercise invalid dimensions, unsupported options, moderation outcomes, network failures, rate limits, timeouts, decoding errors, storage failures, retries, cancellation, and duplicate-job protection.

8. Measure cost per approved asset

   Log model, text and image inputs, quality, dimensions, token usage, retries, latency, reviewer decision, and rejection reason so pricing reflects useful outputs rather than raw requests.

9. Release behind a controlled gate

   Use a feature flag or staged rollout, compare error and approval rates, keep a rollback path for application code, and do not route traffic to the retired DALL-E 3 model.

10. Correct public documentation

    Remove active DALL-E 3 access and price claims, label historical tutorials, publish the May 12, 2026 retirement date, identify the replacement, and schedule future model-lifecycle reviews.

DALL-E 3 Gotchas and Limits to Know Before You Start

• DALL-E 3 API access ended on May 12, 2026.
• Old ChatGPT subscription descriptions should not be presented as current DALL-E 3 pricing.
• Historical per-image API prices no longer describe an available product.
• GPT Image 2 pricing is token based and varies with the complete image job.
• DALL-E 3 quality values do not map directly to current GPT Image quality values.
• The vivid and natural style options were specific to DALL-E 3.
• Old code may expect temporary URLs while GPT Image outputs require different handling.
• DALL-E 3 generated one image per request and did not provide the current editing workflow.
• Revised prompts can make archived results difficult to interpret when only the original prompt was stored.
• GPT Image 2 supports broader dimensions, but requests must follow documented resolution and aspect-ratio limits.
• Model access can require organization verification and is subject to account rate limits.
• A successful model-name change does not prove quality, safety, storage, or cost compatibility.

Which DALL-E 3 Feature Fits Your Use Case

Starter Prompts for DALL-E 3

Is DALL-E 3 still available through the OpenAI API?

No. OpenAI removed the dall-e-3 model from the API on May 12, 2026 and marks it as deprecated and unavailable.

When was DALL-E 3 removed?

OpenAI announced the deprecation in November 2025 and shut down API access to dall-e-3 on May 12, 2026.

What replaces DALL-E 3?

OpenAI recommends GPT Image 2 for current image generation and editing. Its retirement notice also named GPT Image 1 and GPT Image 1 Mini as replacements.

Should a new application use DALL-E 3?

No. The model is unavailable. New OpenAI image integrations should evaluate GPT Image 2 and follow the current Image API or Responses API guidance.

Does old DALL-E 3 pricing still apply?

No. Historical per-image prices no longer describe an available service. GPT Image 2 uses token-based pricing for text input, image input, cached input, and image output.

How was DALL-E 3 different from GPT Image 2?

DALL-E 3 accepted text and generated one image. GPT Image 2 accepts text and image inputs, supports generation and editing, offers broader size and output controls, and reports token usage.

What sizes did DALL-E 3 support?

The API supported 1024 by 1024, 1792 by 1024, and 1024 by 1792 pixel outputs.

What quality settings did DALL-E 3 use?

DALL-E 3 used standard and HD quality settings. Current GPT Image models use low, medium, high, or automatic quality where supported.

What were the vivid and natural settings?

They were DALL-E 3 style controls. Vivid favoured dramatic results, while natural aimed for a less hyper-real treatment. They are not general GPT Image 2 parameters.

What was the DALL-E 3 revised prompt?

DALL-E 3 could return the rewritten prompt used for generation. Preserve it with the original prompt when archiving or evaluating old outputs.

Why can an old DALL-E 3 integration fail after a model swap?

Current models use different parameters, response data, output formats, cost accounting, and image-input capabilities. Updating only the model name is not a complete migration.

How should teams validate a migration?

Use representative production prompts and references, define acceptance criteria, compare quality, safety, latency, file handling, and cost, then test errors, limits, moderation, and storage.

Related Design & Creative Tools

Explore the Network