Two Retirements, One Brand Gone

DALL-E didn’t die in a single event — it was unwound in stages over nine months. The sequence matters for understanding what OpenAI is actually doing with image generation.

The first cut came on November 14, 2025, when OpenAI posted a deprecation notice: dall-e-2 and dall-e-3, the API models that had powered hundreds of applications since 2022, would stop accepting requests on May 12, 2026. That deadline passed this spring with little fanfare. The replacements: three new endpoints — gpt-image-1, gpt-image-1-mini, and gpt-image-2.

The second cut was more visible. On August 30, 2026, OpenAI retired the official DALL-E GPT — one of the original custom GPTs launched alongside the ChatGPT Plugin Store in 2023. Users were warned weeks in advance to download any images stored there before the deadline, since OpenAI made no guarantee of preserving access afterward. The DALL-E GPT is now gone.

Together, these two events mark the end of DALL-E as a product line. The brand that once defined consumer AI image generation has been retired.

GPT Image 2: What You Actually Get

The API replacement, gpt-image-2, launched on April 21, 2026 to the OpenAI API, followed by ChatGPT integration on April 22. The upgrade over DALL-E 3 is genuine, starting at the architecture level.

DALL-E 3 was bolted onto ChatGPT as an external tool — the language model interpreted your prompt and handed it off to a separate image model. GPT Image 2 is native to the GPT architecture. The model reasons about a prompt before generating, which translates to substantially better instruction-following on complex, compositional requests. That architectural shift is the single biggest difference.

The concrete improvements are measurable. Text rendering in images — signs, labels, buttons, code snippets, non-Latin scripts — now works reliably. That was a consistent weak point in DALL-E 3. Output goes up to 2K resolution, and you can request up to eight coherent images per prompt. Photorealism, facial rendering, and UI/screenshot generation are all materially better than the previous generation.

For the majority of ChatGPT users, the transition was invisible. ChatGPT had already migrated to GPT Image 1.5 in December 2025. By the time DALL-E was officially retired in August, most users had been on the new system for eight months without noticing. The August 30 retirement removed the last visible artifact of the old brand — a standalone GPT that most users had already stopped using.

What Developers Actually Lost

The May 12 API shutdown was more disruptive than the ChatGPT-side retirement, and some of that disruption was permanent.

The migration itself was largely mechanical: update your model parameter from dall-e-3 to gpt-image-1 or gpt-image-2, adjust pricing calculations from per-image to token-based billing, and test your outputs. Most applications survived this with a day or two of engineering work.

Two capabilities disappeared without direct equivalents. DALL-E 3’s dedicated inpainting and outpainting endpoints were used in production workflows for precise regional image editing — masking specific parts of an image and regenerating only that region. GPT Image 2 supports editing, but through a different interface that doesn’t map cleanly to existing integrations. Teams that had built inpainting workflows had to rearchitect, not just swap a parameter.

The pricing model change also caught teams off guard. DALL-E 3 billed at a fixed rate per image. The GPT Image models use token-based pricing that varies by resolution and quality level. High-volume image generation pipelines had to remodel their cost structures entirely. Some teams found the new pricing favorable at scale; others saw costs increase for standard-resolution output.

Developers who built custom GPTs using image generation are largely unaffected. Custom GPTs with image capabilities continue to work normally — they route through the GPT Image stack automatically. Only OpenAI’s own official DALL-E GPT was retired on August 30.

If you’re still seeing generation errors and haven’t updated your integration, the DALL-E 2 and DALL-E 3 endpoints have been down since May 12. The migration documentation is on OpenAI’s deprecations page, and gpt-image-1 is the closest drop-in replacement for most workflows.

What the Consolidation Signals

OpenAI’s image strategy now follows the same playbook it has applied to voice, code execution, and web search: absorb the capability natively into the model rather than maintaining it as a separate service. That pattern is consistent and deliberate.

This has consequences for the competitive landscape beyond image generation. Standalone image APIs — whether proprietary or open-source — compete against a moving target that keeps absorbing their most common use cases. Open-source alternatives like Flux and Stable Diffusion 3.5 have gained enterprise traction partly because they offer a stable API surface not subject to OpenAI’s product roadmap decisions. The DALL-E deprecation gave that argument more credibility: if you’ve built a production workflow on an OpenAI-branded product, plan for it to be consolidated or retired within a few years.

The super-app strategy — folding browsing, image generation, code execution, voice, and memory into a single ChatGPT interface — is also a deliberate moat-building move. Each integrated capability reduces the need for third-party tools and increases switching costs. That’s a legitimate strategy, and OpenAI is executing it consistently. But it creates real dependency risk for teams that have standardized on a specific product surface. For more on how OpenAI’s strategy has evolved at a corporate level, our analysis of their S-1 filing covers the structural bets they’re making.

The one area where the new stack is unambiguously better: instruction-following. If your use case involves generating images with specific text, precise layout requirements, or multi-element compositional prompts, GPT Image 2 is a real upgrade. The reasoning-first architecture closes a gap that DALL-E 3 never solved. That’s worth knowing before you dismiss the migration as just a branding reshuffle — it isn’t.

What the DALL-E retirement ultimately confirms is that evaluating any AI tool should include an honest assessment of the vendor’s track record with deprecation. Nine months between announcement and shutdown is reasonable notice. But teams that built deeply on DALL-E’s specific inpainting API found out that the replacement wasn’t a drop-in. That’s the pattern to watch for as OpenAI continues consolidating more capabilities into the core platform.

Further Reading

• OpenAI Deprecations Page — Official timeline and migration paths for every deprecated model, including all DALL-E variants.
• DALL-E Is Dead: What Replaces It — Genra.ai — Detailed breakdown of the May 12 API shutdown and what each GPT Image model does differently.
• What is GPT Image 2? — MindStudio — Technical specs, architecture differences, and a comparison with DALL-E 3 for developers evaluating the migration.