European Alternatives to DALL-E

Why people leave DALL-E

DALL-E was the name OpenAI gave its first mainstream text-to-image models, and for many people it is still the word for "type a sentence, get a picture". The name now describes something that has been taken down.

OpenAI's deprecation page lists dall-e-2 and dall-e-3 with a shutdown date of 12 May 2026, and the dall-e-3 model page says in plain words that the model has been deprecated and removed from the API. Anyone who built on it has already been moved.

The replacement is the GPT Image family, and that changes what you are buying.

Instead of a fixed price per picture, gpt-image-2 is billed per million tokens: according to OpenAI's pricing page, $8.00 for image input and $30.00 for image output, with text input at $5.00. A single image therefore has no single price, and the cost of a campaign depends on sizes, quality settings and how many edit rounds it takes. OpenAI offers a calculator in its image guide for exactly this reason.

People looking beyond it rarely complain about image quality. They complain about the surrounding terms: a model name that disappears from under a production workflow, token-based invoices that are hard to forecast, an organisation verification step before some image models can be used, and prompts and uploads that are stored with a US provider for abuse monitoring. The European tools below answer those worries in different ways, and some of them do not try to match the model at all.

• The model you built on was retired OpenAI removed dall-e-2 and dall-e-3 from its API on 12 May 2026 and points developers to gpt-image-2, gpt-image-1 or gpt-image-1-mini. Prompts tuned for one model rarely behave identically on the next. Open-weight models from Black Forest Labs and Stability AI cannot be switched off by a vendor, because the files sit on your own disk, which turns a deprecation notice into an upgrade decision you take yourself.
• Image pricing is per token, not per picture On OpenAI's pricing page, gpt-image-2 lists image input at $8.00, cached input at $2.00 and image output at $30.00 per million tokens, plus $5.00 per million for text input. A finance team cannot multiply that by a number of pictures without first modelling image sizes and quality. Freepik, Pincel and Neuroflash sell a plan or a credit allowance instead, which is easier to budget even when it is not cheaper.
• Storage and retention follow OpenAI's defaults For the image generation and image edit endpoints, OpenAI's data controls page states that data is not used for training and that abuse-monitoring logs are kept for up to 30 days. Zero Data Retention is available only with limitations. A storage region outside the United States needs sales approval and a Modified Retention amendment, and system data may still be processed elsewhere.
• Access can depend on verification OpenAI's image guide says you may need to complete API Organization Verification in the developer console before using GPT Image models. For a small agency or a freelancer building a quick prototype, that is one more gate between an idea and a first image. Self-hosted open weights have no such step, because nothing is called remotely.

What you have to replace, not just match

Replacing DALL-E means replacing a prompt-to-image habit, not a product. Most teams use it in one of three ways: through a chat window where each follow-up message refines the picture, through the image API inside their own software, or as a source of illustrations for decks and campaigns.

Each of those has a different European answer. Chat-style refinement is closest in Neuroflash, API use is closest in Black Forest Labs and Stability AI, and a mixed pipeline of generation plus a licensed stock fallback is what Freepik is built around.

The second thing to replace is everything that surrounds the generation call: the moderation behaviour you have learned to expect, the image sizes your layouts assume, the way an edit request keeps most of a picture intact, and the invoices your finance team already reconciles.

Inpainting, the editing function the gpt-image-2 model page lists, has no single European counterpart. PhotoRoom, Pincel and Luminar Neo each cover a slice of it, which is why this page lists them lower down as editing tools rather than as generators.

Finally, check what your generated images are used for. A decorative blog illustration, a product shot that must resemble a real item and an ad that carries your brand colours are three different requirements. DALL-E users who only ever needed the first can pick almost anything on this list. Those who needed the second or third will find that a tool built around a real photograph or a defined style does more for them than a larger prompt window.

The alternatives compared

How each alternative compares to DALL-E

What actually breaks when you switch

The first thing to break is the prompt library. A sentence that gave you a clean product shot from a GPT Image model may give a different composition from FLUX or a Stable Diffusion variant. Set aside a test day: run your ten most used prompts through two candidates, keep the outputs side by side, and let the people who use the images pick, not the person who wrote the prompts.

The second break is plumbing. If your code calls the OpenAI image endpoints, you will change the request format, the size options and the way errors and moderation blocks come back.

Retention also differs: OpenAI states 30-day abuse-monitoring logs on those endpoints, and your new vendor will have its own figure, so update your privacy notice. Keep your OpenAI account open until the new route has produced a full month of work, because billing history and past generations are easier to retrieve while the account exists.

Is DALL-E still available, and what do I switch from?

As an API model it is not. OpenAI's deprecation table gives 12 May 2026 as the shutdown date for dall-e-2 and dall-e-3, and its model page for dall-e-3 says the model is deprecated and removed. The current route in OpenAI's own documentation is the GPT Image family, which accepts text and image input and supports image generation, image edits and inpainting.

So when you compare a European tool with DALL-E, you are really comparing it with whichever GPT Image model your team now calls. Write down which one, because the token price, the image sizes and the edit behaviour differ per model and you want a like-for-like test with the same ten prompts.

Can I get a flat price per image in Europe?

Sometimes. Freepik sells a subscription with a free tier that requires attribution, Pincel sells monthly plans with a credit allowance that carries over while the plan stays active, and Neuroflash sells seats at a published monthly price billed annually. Black Forest Labs bills its API by usage, and its open-weight models cost nothing to run except the hardware.

None of these gives you a universal price per picture, because credits, plan tiers and model choice all move the number. What they do give is a ceiling you can put in a budget, which a per-token invoice does not.

Which one to pick

If you call an image API from your own software, start with Black Forest Labs, and add Stability AI if you want to fine-tune on your own images or must run everything on hardware you control. If your team works in a chat window and makes visuals for campaigns, Neuroflash is the closest working pattern, and Freepik is the better pick when a licence record per asset matters more than the last bit of image quality.

If your real job is editing photographs, you were probably using the wrong tool in DALL-E all along. PhotoRoom suits product catalogues, Pincel suits occasional edits paid by credit, Luminar Neo suits photographers who want a one-time licence, and Nero AI fits as the enlarging step. Stay with OpenAI if you need its exact model behaviour and your data rules allow a US provider; none of the tools above is a drop-in copy.