Midjourney vs DALL-E vs Stable Diffusion: Picking an AI Image Generator

“Midjourney vs DALL-E vs Stable Diffusion” is still how most people type this search, even though one of those three names barely exists as a product anymore. That’s worth clearing up before anything else, because it changes what’s actually being compared: this isn’t three symmetric competitors sitting side by side. It’s a subscription-only art tool, a feature now folded into a chat app, and an open-weight model you can run yourself. All three turn a text prompt into an image. How you reach each one, what it costs, who owns the result, and how much control you get over the output are different enough that “which is best” is close to the wrong question. Here’s what each one actually is as of September 2026, what it costs right now, and which one fits your situation.

Midjourney: the one with the house style

Midjourney is a paid-only image generator you now reach mainly through a web app at midjourney.com — Discord, which used to be the only way in, is now an optional sign-in method rather than a requirement, and most users work from the browser interface today [1][3]. There has never been a free tier since Midjourney removed its original trial in March 2023, and that hasn’t changed [1].

The current default model is V8.1 (default since 11 June 2026), with V8.2 following in July 2026 as a refinement pass — bolder aesthetics, fewer weak outputs, better personalization [2]. What hasn’t changed across versions is the reason people specifically ask for Midjourney by name: its default output looks more like finished art — painterly, cinematic, considered — than a typical diffusion model’s default, without any prompt-engineering effort. That’s a real, distinctive product quality, not marketing.

Pricing is a flat monthly tier, no per-image metering, paid in USD, with roughly 20% off for annual billing:

“Fast” hours generate immediately; once you run out, Standard and above switch to unlimited but slower “Relax” generation rather than cutting you off [1]. All paid tiers give you commercial usage rights over what you generate — but if your business earns more than $1 million a year in gross revenue, Midjourney’s terms require you to be on the Pro plan specifically to keep those commercial rights, regardless of which tier you’d otherwise choose [7]. There is no first-party API: anything programmatic runs through unofficial, unsupported wrappers, which is a real gap if you’re building a product rather than making pictures for yourself.

DALL-E doesn’t really exist anymore — meet GPT Image, living inside ChatGPT

This is the fact most in need of updating if what you remember is “DALL-E 3.” It’s retired. OpenAI deprecated both dall-e-2 and dall-e-3 in its API on 12 May 2026, and separately retired the standalone “DALL·E” app inside the ChatGPT store on 30 August 2026 [4][8]. Neither exists as a live product anymore.

What replaced it is GPT Image 2 (gpt-image-2), which reached general availability on 21 April 2026 and is now OpenAI’s default image-generation and editing model, both in the API and inside ChatGPT itself, where it’s branded simply “ChatGPT Images” [4][5][9]. That naming is the real change to understand: image generation is no longer a separate destination or brand, it’s a capability of the chat interface you’re already in — you type “make me a poster of a fox reading in a library,” and it generates inline, in the same thread, editable by follow-up messages (“make the sky darker”) rather than by re-writing the whole prompt.

That’s also its biggest practical advantage: onboarding is trivial. If you already have a ChatGPT account — including the free tier, which now includes ChatGPT Images with usage caps and queuing rather than gating it behind a paid plan entirely — you already have access [8][9]. ChatGPT Plus ($20/month) and Pro (OpenAI’s higher tier, priced at $200/month) raise the generation limits substantially; see AI subscription pricing 2026 for this wiki’s tracked, current figures across all the major chat subscriptions, since exact regional pricing shifts more often than this page should try to re-verify.

Via the API, pricing is token-based rather than a flat per-image fee, which is a genuine mechanical difference from how DALL-E used to bill: OpenAI’s own pricing documentation lists $5 per million text-prompt tokens, $8 per million image-input tokens (for edits), and $30 per million output tokens for gpt-image-2, batch processing at roughly half that [6]. In practice that works out to roughly $0.02–$0.20 for a single image depending on resolution and quality setting — check OpenAI’s own cost calculator before budgeting a specific volume, because the exact figure moves with size and quality far more than a flat-fee competitor’s pricing does [6].

The tradeoff for that convenience: GPT Image’s default aesthetic is generally more literal and less immediately “artistic” than Midjourney’s out of the box, and because it inherits ChatGPT’s content policy, it’s noticeably more conservative than either Midjourney or an uncensored local Stable Diffusion checkpoint on things like real people’s likenesses, some copyrighted characters, and other edge-case requests — a real constraint if that’s specifically what you need.

Stable Diffusion: the only one you can actually own

Stable Diffusion is the odd one out in the most consequential way: it’s open-weight, published by Stability AI, and you can download and run it yourself rather than renting access to someone else’s server. The current flagship generation, confirmed directly against Stability AI’s own site, is still Stable Diffusion 3.5 — Large, Large Turbo, and Medium variants, built on a Multimodal Diffusion Transformer architecture [10][11]. A number of low-quality aggregator sites currently circulate claims of a “Stable Diffusion 4” having launched in April 2026; Stability AI’s own news feed shows no such announcement as of its most recent entries (a Series B funding round on 25 August 2026), and its product page continues to describe 3.5 as “our most powerful image model yet” [10][11] — treat any “SD4” claim you’ve seen elsewhere as unverified until Stability AI’s own site says otherwise.

You run it two ways. Locally, using ComfyUI (node-based, more control) or the Automatic1111 WebUI, on a consumer GPU with roughly 6GB+ VRAM or an Apple Silicon Mac — after that hardware cost, generation is free, with no per-image bill, ever. Hosted, through the Stability AI API, Replicate, Hugging Face Inference, or Amazon Bedrock, at per-image pricing if you’d rather not manage a GPU — see Stable Diffusion on this wiki for current API rates and setup detail. Licensing matters here in a way it doesn’t for the other two: Stability AI’s Community License is free for commercial use as long as your organization earns under $1 million a year in revenue; above that threshold, you need a paid Enterprise License [12].

What actually distinguishes Stable Diffusion from the other two is the ecosystem built on top of the open weights, not the base model’s default output — which, unfine-tuned, generally trails Midjourney’s and GPT Image’s out-of-the-box quality. LoRA and DreamBooth let you fine-tune a custom style or a specific subject on a few dozen images in under two hours on a single GPU; ControlNet lets you constrain a generation to a specific pose, depth map, or edge layout rather than hoping the prompt gets it right; and Civitai hosts thousands of community-trained checkpoints and LoRAs for free. That combination — fine-tuning plus spatial control plus zero marginal cost — is not something either of the other two products offers at all, at any price. The real cost is technical: getting a local install running, understanding sampler settings and CFG scale, and building a ComfyUI workflow all require a learning curve neither Midjourney nor ChatGPT asks of you.

Side by side

Which one should you actually pick

You want the best-looking images with the least effort, and you don’t mind a monthly bill: Midjourney. Nothing else produces that finished, art-directed look by default, and V8.1/V8.2 closed most of the old gap in following detailed instructions.

You already pay for ChatGPT, or you want the simplest possible starting point: GPT Image, inside ChatGPT. If you’ve never generated an AI image before, this is genuinely the lowest-friction way to try it — free tier included — and the conversational editing (“now make it nighttime”) beats re-prompting from scratch in every other tool here.

You want zero ongoing per-image cost, full creative control, or a specific consistent style across hundreds of images: Stable Diffusion, run locally. This is also your only option if privacy or data residency rules mean the images can’t touch someone else’s server at all.

You’re building a product that generates images programmatically: Stable Diffusion (self-hosted or via a hosted API) or the GPT Image API. Not Midjourney — it still has no supported first-party API, full stop, which rules it out for this use case regardless of how good its output looks.

You run a business over $1M/year in revenue and want to use Midjourney commercially: you specifically need the Pro plan ($60/month), not Basic or Standard — check this before you subscribe, not after.

None of this is permanent. All three vendors ship new model versions every few months, and OpenAI has already renamed and re-architected this product category once in the past year. Re-check current pricing and model names directly before budgeting a project around any of these figures.

Sources

1. Midjourney, official pricing (Basic/Standard/Pro/Mega tiers, fast-hour allocations, Relax and Stealth mode availability, no free tier since March 2023): https://www.midjourney.com/pricing — cross-verified against independent third-party pricing trackers, fetched September 2026.
2. Midjourney, official updates blog on V8.1 becoming the default model (11 June 2026): https://updates.midjourney.com/v8-1-is-now-the-default-model/
3. Dexerto, “Midjourney drops Discord requirement for AI image generation,” on the web app becoming the primary interface with Discord as an optional login: https://www.dexerto.com/tech/midjourney-drops-discord-requirement-for-ai-image-generation-2874138/
4. OpenAI, API deprecations documentation, confirming dall-e-2/dall-e-3 shutdown 12 May 2026 and gpt-image-2 as the recommended replacement, fetched September 2026: https://developers.openai.com/api/docs/deprecations
5. OpenAI, GPT-Image-2 model documentation, fetched September 2026: https://developers.openai.com/api/docs/models/gpt-image-2
6. OpenAI, API pricing documentation (image-model token rates), fetched September 2026: https://developers.openai.com/api/docs/pricing
7. terms.law, Midjourney commercial-use policy summary, on the $1M revenue / Pro-plan requirement for retaining commercial rights: https://terms.law/ai-output-rights/midjourney/
8. NotebookCheck, “DALL-E leaves ChatGPT on August 30,” on the retirement of the standalone DALL·E GPT and its replacement by ChatGPT Images: https://www.notebookcheck.net/DALL-E-leaves-ChatGPT-on-August-30-download-your-images-first.1360522.0.html
9. This wiki, OpenAI API , independently confirming the same DALL-E retirement and GPT Image 2 timeline from OpenAI’s own documentation.
10. Stability AI, Stable Image product page, confirming Stable Diffusion 3.5 as the current flagship release, fetched September 2026: https://stability.ai/stable-image
11. Stability AI, News & Updates feed, showing no Stable Diffusion 4 announcement through the most recent entry (25 August 2026), fetched September 2026: https://stability.ai/news-updates
12. Stability AI, Community License terms ($1M annual revenue threshold for free commercial use): https://stability.ai/license

Further reading

• Stable Diffusion : setup, code samples, ControlNet and LoRA workflows, and current hosted-API pricing for the one tool in this comparison you can self-host.
• OpenAI API : the API side of GPT Image, alongside OpenAI’s language models.
• AI subscription pricing 2026 : current ChatGPT, Claude, and Gemini consumer tier pricing, tracked separately from the API rates above.
• What is generative AI? : the category all three of these products sit inside.
• Diffusion models : how the underlying technology behind all three actually works.
• LoRA : the fine-tuning technique that gives Stable Diffusion its style-customization edge.