DALL·E 3 Is Retired. Here's How to Prompt GPT Image 2.

OpenAI shut down DALL·E 3 on May 12, 2026. Its successor — now GPT Image 2.5, the default across ChatGPT — prompts differently: it plans before it draws. What changed, what carries over, and how to get consistent results across a batch.

On May 12, 2026, OpenAI removed DALL·E 2 and DALL·E 3 from its API. If you had prompts, workflows, or muscle memory built around DALL·E 3, they now point at a model that no longer exists. The replacement is GPT Image 2 (API id: gpt-image-2), released April 21, 2026 — and it is not just DALL·E 4 with a new name.

What actually changed

GPT Image 2 is the first mainstream image model that *plans before it draws*. It reasons about layout, composition, and text placement before committing pixels, and it self-checks its output. In practice that means:

Complex instructions land. Multi-object scenes with spatial relationships ("the cat is behind the window, reflected in the mirror") stopped being a coin flip.
Text in images is reliable. Signs, labels, UI copy — legible on the first try far more often.
4K output is in beta. The 1024px ceiling is gone.

What carries over from DALL·E 3 prompting

The descriptive-sentence style still works. GPT Image 2 wants natural language, not tag soup — the opposite of classic Stable Diffusion prompting. Front-load the subject, describe the scene the way you'd brief a photographer, and put constraints ("no text, no watermark") in a separate sentence.

What to stop doing

Stop over-hedging. DALL·E 3 rewrote your prompt behind the scenes, so people learned to repeat key ideas three ways to survive the rewrite. GPT Image 2 follows instructions closely enough that repetition now causes overemphasis — you get three cats because you mentioned the cat three times.

Updated October 2026: GPT Image 2.5

The line moved on. GPT Image 2.5 is now the default image model across every ChatGPT tier, and it sits at the top of the blind-vote image leaderboards. The prompting approach below is unchanged — 2.5 is a sharper, higher-fidelity version of the same plan-then-draw model — but two things got better: 4K is standard, and character and layout stay consistent across generations far more reliably.

How Prompt Architect handles it

The Studio's GPT Image slot generates the descriptive-prefix format the model likes, with your negative constraints and aspect ratio appended as clean sentences. Old share cards and remix QR codes that pointed at DALL·E 3 automatically load with GPT Image selected — your saved recipes are not orphaned.

The real win now isn't one prompt — it's *repeatability*. Save your subject framing, palette, lighting and negatives as a reusable Style System, apply it in a click, and your tenth image matches your first instead of drifting. That's the difference between a lucky generation and an on-brand set.

Want to see how the same idea renders across GPT Image 2.5, Midjourney V8.2, FLUX.2, and Nano Banana Pro? That's what Compare view is for.

— The Prompt Architect team