DALL·E Is Gone: What to Use Instead in 2026

TL;DR — DALL·E is gone, here is what to do

Short on time? Here is the gist. Every point is unpacked below with dates, prices and sources.

• On August 30, 2026 OpenAI retired the official DALL·E GPT in ChatGPT — the last surface where the DALL·E name still existed for ordinary users.
• The dall-e-2 and dall-e-3 API endpoints were shut down earlier, on May 12, 2026. Developers were notified on November 14, 2025 — six months of notice.
• Save the images from your old DALL·E GPT conversations first: OpenAI explicitly advises downloading anything you want to keep before the tool disappears.
• Image generation inside ChatGPT itself did not go anywhere — it has been running on GPT Image, not DALL·E, for a while. User-created GPTs with image generation enabled are unaffected.
• OpenAI’s official replacement is gpt-image-2, plus gpt-image-1 and gpt-image-1-mini for cheap high-volume work.
• The real choice is wider than one model: GPT Image 2 for prompt adherence, Midjourney for art, FLUX.2 for photorealism, Nano Banana Pro for editing an existing photo, Ideogram 3 for text inside the image.
• In Twin AI GPT Image 2, Flux 2 Pro and Nano Banana Pro run on one credit balance (Midjourney and Ideogram 3 are not in Twin AI — Flux 2 Flex and GPT Image 2 cover their jobs), with no VPN and local-card payment — start at /create/photo.

What exactly was shut down, and when: the full timeline

The confusion around “DALL·E is closed” exists because it was not closed once — it was closed in four steps over eighteen months. Plenty of people still believe they are drawing in DALL·E while in fact they have been using a different model for a year. Here is the sequence, so you know what stopped working and when.

• March 2025 — image generation in ChatGPT moves to the native multimodal GPT Image architecture. DALL·E 3 formally stays, but stops being what draws by default.
• December 2025 — gpt-image-1.5 and gpt-image-1-mini ship in the API, markedly stronger than their predecessors. On November 14, 2025 OpenAI sends developers the formal DALL·E deprecation notice.
• May 12, 2026 — dall-e-2 and dall-e-3 are removed from the API. Requests to those snapshots no longer work; the documented replacements are gpt-image-2, gpt-image-1 and gpt-image-1-mini.
• August 30, 2026 — the official DALL·E GPT in ChatGPT is retired. The DALL·E name leaves OpenAI’s products for good.

What to do right now if you used DALL·E

First and urgent: the images. OpenAI specifically warned that anything you generated inside the DALL·E GPT and want to keep should be downloaded. Once the GPT itself is gone there is no convenient way back into those conversations, and nobody will restore them after the fact. Walk your history, export what matters into a folder — it is twenty minutes of work.

Second: work out what actually broke for you. If you are a regular ChatGPT user who just asks “draw me a picture” in the normal chat, nothing broke — GPT Image has been doing that for a while and it is still there. If you built your own GPT with image generation enabled, it keeps working too. Two groups really got hit: people who opened the DALL·E GPT as a separate tool, and developers whose code already hit dead dall-e-2 and dall-e-3 endpoints back in May.

Third: pick what you draw with next. This is worth half an hour rather than grabbing the first option — in the three years since DALL·E’s peak the market changed beyond recognition, and the official “heir” is not automatically the best fit for your task. The model breakdown is in the next section and on /alternatives/dall-e.

What to use instead of DALL·E: five models for 2026

The direct heir on the OpenAI line is GPT Image 2. Per the Artificial Analysis Image Arena it holds first place in text-to-image as of August 2026 with an Elo around 1370 — the largest first-to-second gap that leaderboard has recorded. But “first on average” does not mean “best for your job”: each of the five below owns a zone where it wins.

• GPT Image 2 (OpenAI) — the direct DALL·E replacement: best at long, complex prompts, holds scene logic, renders native 2K against DALL·E 3’s 1024 px ceiling and puts legible text on an image far more reliably. Breakdown at /alternatives/gpt-image-2.
• Midjourney — the benchmark for artistic quality: art, concepts, stylisation, a striking frame straight out of the box. Where DALL·E was correct but bland. See /alternatives/midjourney.
• FLUX.2 Pro (Black Forest Labs) — photorealism, anatomy and difficult lighting; the pick for product shots and high-volume API work. See /alternatives/flux-2-pro.
• Nano Banana Pro (Google) — the best instruction-driven editor: swaps an object, background or detail on an existing photo without touching the rest, and pushes resolution to 4096×4096. In the editing arena it sits right behind GPT Image 2. See /alternatives/nano-banana-pro.
• Ideogram 3 — typography: logos, posters and banners with readable lettering. DALL·E’s historical weak spot is closed exactly here. See /alternatives/ideogram-3.

One prompt, three styles — different models read “draw a bicycle” differently

Which one to pick for which job

A short mapping so you do not have to guess. Precise adherence to a long brief, infographics, collages, schematics — GPT Image 2. Art, covers, concepts, anything where aesthetics lead — Midjourney. Honest product shots and portraits with correct anatomy and skin — FLUX.2 Pro. Editing an existing photo from a text instruction (replace the background, remove an object, change a colour) — Nano Banana Pro. A logo, poster or banner with readable text — Ideogram 3.

When the task sits on a border, it is safer to test than to guess. Twin AI has /compare for exactly that: one prompt goes to several models at once, the results arrive side by side, and you pick the winner by eye. That is the same check that in ChatGPT would mean paying for a separate subscription per service.

For developers: migrating code from dall-e-3 to gpt-image-2

If your bot or backend still calls dall-e-3, it broke back on May 12, 2026 — the endpoint returns an error, not an image. The migration is simple, but a couple of things genuinely change:

• Swap the model name: dall-e-3 → gpt-image-2 (or gpt-image-1 / gpt-image-1-mini if price matters more than quality). That is exactly what the deprecations documentation recommends.
• Check how you read the response. The GPT Image family returns the image differently from DALL·E — code that blindly grabs the first URL may stop working.
• Revisit your prompts. GPT Image follows long instructions far better, so DALL·E-era crutches (“highly detailed, 8k, masterpiece”) are unnecessary and often actively hurt.
• Budget for different pricing and sizes: native 2K instead of 1024 px, with per-image cost depending on quality and resolution — recalculate against your real traffic profile, not the old DALL·E rate.
• If you need one key across several vendors’ models at once, that is what Twin AI does: one account and one balance covering GPT Image 2, Flux 2 Pro, Flux 2 Flex, Nano Banana Pro and the rest (Midjourney is not in Twin AI) — see /models.

What image generation costs now that DALL·E is gone

Inside the OpenAI ecosystem images are drawn by GPT Image. Through ChatGPT that means a Plus subscription at $20/month with a soft cap of roughly 50 images per three hours. Through the API you pay per image, and the price depends on model, quality and resolution: from a few cents on gpt-image-1-mini at low quality up to noticeably more on the flagship at 2K. Either way payment runs through a foreign gateway, which is a separate problem for users in Russia.

Twin AI works differently: pay-as-you-go credits instead of a subscription. An image costs 200 credits (≈$1.20), you pay only for generations you actually run, and the same balance covers photo, video and chat. The free credits you get at signup cover 10–15 generations to try it, with no card required. Payment goes through local cards and SBP. Current rates are on /pricing.

How Twin AI answers this in one window

The real annoyance of leaving DALL·E is that the five models above live in five different services: one is subscription-only, another needs an API key, a third does not open from Russia, a fourth declines the card at checkout. Assembling a workflow out of five tabs and a VPN is expensive in both time and money.

Twin AI is a single app — web and iOS — where GPT Image 2, Flux 2 Pro, Flux 2 Flex, Nano Banana Pro, Seedream and other models run on one account and one credit balance (Midjourney and Ideogram 3 are not in Twin AI — Flux 2 Flex and GPT Image 2 stand in for them). Two things ChatGPT does not have: the /compare mode, which sends one prompt to several models at once and shows the results side by side, and the Twin Elo leaderboard at /models, built from hundreds of thousands of blind A/B votes by real users — it tells you which model usually wins your kind of task before you spend a credit. Everything works from Russia without a VPN, and generations are private by default.

Common mistakes when migrating off DALL·E

What usually goes wrong for people who move in a hurry:

• Not downloading the images from the DALL·E GPT before the shutdown — there is nowhere to recover them from afterwards.
• Hunting for an exact DALL·E clone. There is not one: GPT Image is a different architecture with different editing behaviour, and re-learning it beats searching for a copy.
• Porting old prompts word for word. Tags like “8k, masterpiece, highly detailed” were written around DALL·E’s weak instruction-following and tend to hurt on modern models.
• Picking one model for everything. A logo, a portrait and an edit of an existing photo are won by different models — running the prompt through /compare is faster than arguing about taste.
• Forgetting access and payment. A model can be excellent and still useless if the service needs a VPN and declines your card.