This comparison needs a correction before it needs an update, so here it is: one of the two tools in the title is on its way out.
OpenAI announced the DALL-E GPT would be retired from ChatGPT on August 30, 2026. When I checked on August 31 the GPT page still resolved, with OpenAI describing DALL-E there as its legacy image generation model and pointing users to the main chat for the current one, so treat it as on the way out rather than confirmed gone. The DALL-E API was shut down earlier and separately, on May 12, 2026, which is not in doubt. OpenAI's retirement notice tells users to download any images they want to keep, and that advice stands regardless of exactly when the entry point disappears.
If you came here to decide between Midjourney and DALL-E, that decision has partly been made for you.
What you should do this week
Two things, and the first one has a deadline.
If you have images sitting in DALL-E conversations that you would be annoyed to lose, export them now. OpenAI is explicitly telling people to download what they want to keep. I would not spend energy arguing about whether that means the files vanish on the 31st. When a vendor tells you to save your work, save your work.
The second thing is less urgent: stop treating "DALL-E" as the name of the thing you buy. Image generation inside ChatGPT continues. It runs on ChatGPT Images 2.0, which OpenAI says is available on all ChatGPT plans, with better text rendering, multilingual support, and visual reasoning. There is also a mode OpenAI calls images with thinking, where the model plans and refines before it generates rather than answering immediately.
So nothing is being taken away from you except a name and an old entry point. That is worth knowing, because a lot of articles will spend the next month telling you DALL-E died and implying you need to go elsewhere.
What this comparison can and cannot tell you now
The earlier version of this post described two weeks of side by side testing across six prompt categories, with specific hit rates for hands, text accuracy, and prompt adherence.
That testing was real, and it is also out of date on both sides. It ran against DALL-E 3 and against Midjourney around v6.1. Midjourney has shipped V8 in March 2026, V8.1 in April, and V8.2 became the default on July 24, 2026. DALL-E has been replaced entirely.
I am not going to take findings from an older generation of both models and quietly present them as current. That is the single most common thing wrong with AI tool comparisons, and I would rather have a shorter honest post than a detailed dishonest one. So the specific numbers are gone.
What I will do is tell you which parts of the comparison are structural, meaning they come from how the products are built rather than from a particular model version, and are therefore still worth reasoning about.
The structural difference that has not changed
Midjourney is a specialist. It does one job and sells subscriptions for that one job. You reach it through Discord or the web app, you generate variations, you upscale what you like.
Image generation in ChatGPT is a feature of a general assistant you may already be paying for. That is the whole buying logic and it survives every model update. If you already subscribe for writing or coding, the images are effectively bundled. If you do not, you are comparing a dedicated tool against a subscription that happens to include images.
The consequence people underrate: iteration works differently. In a chat you can say "same scene, add a cat on the windowsill" and keep going in conversation. Midjourney has parameters and reference images to steer output, which give you more precise control once you learn them and less natural control before you do. Neither approach is better in the abstract. One suits people who think in conversation, the other suits people who think in settings.
The tradeoff everyone repeats, and how much to trust it
The received wisdom is that Midjourney wins on aesthetics and OpenAI wins on instruction following and readable text.
That was true when I tested it and I suspect the aesthetic half still is, because it reflects a deliberate product choice rather than a temporary gap. Midjourney tunes for images that look striking. That is what its users buy.
The text half I am much less sure about now, and here is why. Midjourney's weakness was rendering readable words inside an image, and OpenAI's advantage there was large. But improved text rendering is one of the specific things OpenAI claims for ChatGPT Images 2.0, and Midjourney has shipped three versions since I looked. So the gap may have widened, narrowed, or stayed put. I have not measured it and I am not going to guess.
If typography inside the image matters to your work, that is the one thing worth testing yourself before you commit, because it is the axis most likely to have moved.
Exactly what breaks on August 30, and what does not
Worth separating, because the headlines will blur it.
The DALL-E API is already gone. That shut down on May 12, 2026, so anything you built against it has either been migrated or has been broken for months. OpenAI points developers at the gpt-image models instead.
The DALL-E GPT inside ChatGPT is the thing being retired. That is the entry point some people still use out of habit, the one that appears as a named GPT rather than just asking ChatGPT for an image. That specific thing is what stops, whenever it finally does.
Image generation in ChatGPT does not stop. Ask for an image in a normal conversation and you get one, on any plan including the free one. So if your workflow is "open ChatGPT, describe a picture", nothing happens to you on the 30th.
The part that needs action is the images already sitting in old DALL-E conversations. OpenAI's guidance is to download what you want to keep. Do that this week rather than in the last few days, because the practical risk is not that files evaporate on a schedule, it is that you forget, come back in November for a client's thumbnail, and find the route to it has been removed.
Prompt style: the difference that survives version changes
The two tools reward genuinely different writing, and this is worth understanding because it does not reset when a new model ships.
Midjourney responds to atmosphere. Words about mood, light quality, film stock, era, and art style push it somewhere useful. Something like "moody Nordic forest at dusk, mist, cinematic composition" gives you close to what those words suggest, because the model is tuned to interpret aesthetic direction. It also takes parameters, so aspect ratio and how much creative licence to allow are dials rather than sentences.
ChatGPT responds to description. Explain the picture the way you would to someone who cannot see it: who is in it, where they are, what the light is doing, what words should appear. It also expands your prompt internally, which is why you sometimes get details you did not ask for.
Paste a Midjourney prompt into ChatGPT and you get something competent and a bit generic. Paste a ChatGPT prompt into Midjourney and you get something literal and visually flat. People conclude one tool is worse when what actually happened is that they used the other tool's language.
The iteration difference follows from this. In a chat you refine by conversation: same scene, move the subject, change the season. With Midjourney you refine by re-prompting and by parameters, which is more precise once learned and less obvious before.
What each side structurally cannot do
Midjourney does not give you conversational editing of an existing image, and it is not the tool for inpainting or fine local edits the way a Stable Diffusion setup is. It also has no general assistant attached, so it is a subscription that does exactly one thing.
ChatGPT's image generation is more filtered, so moody or edgy prompts get softened or declined more often. It does not offer Midjourney's style reference parameters for steering output from an uploaded image, and it has no equivalent of stealth mode, which matters if you do not want your prompts visible in a public gallery.
If neither fits, the useful alternatives split by what you actually need. Adobe Firefly for commercial safety, because it trained on licensed stock. Ideogram when the image is mostly typography. Stable Diffusion locally when you want total control and are willing to run it yourself. Several free options are covered in best free AI image generators, with the caveat that the free tiers in that category have been moving.
What actually changed about the money
This is the part of the comparison that genuinely shifted, and it has nothing to do with image quality.
ChatGPT's image generation is now available on the free tier. Midjourney has no free tier and has not had one since 2023. A year ago both sides of this comparison cost money and the question was which $20 to spend. Now one side has a $0 option and an $8 option below its $20 plan, and the other starts at $10.
Reported Midjourney pricing is four tiers at $10, $30, $60, and $120 a month, with about 20% off annually, and fast GPU time of roughly 3.3, 15, 30, and 60 hours. Standard at $30 is the tier most working users land on, because relaxed mode generation stops being metered and the Basic allowance disappears faster than you expect once you start iterating.
I need to flag the sourcing on those figures. Midjourney blocks automated access to its own site, so I could not verify them the way I verified the OpenAI retirement dates. They are consistent across several independent trackers, which is worth something and is not the same as reading the vendor's page. Check it before you pay.
For ChatGPT, the tiers are Free, Go at $8, Plus at $20, Pro at $100 and $200, then Business and Enterprise. I went through what each one is worth in the ChatGPT Plus review.
Which one to pick now
If you generate images occasionally and cost matters, start with ChatGPT's free tier. That advice did not exist a year ago and it makes a real difference to people who were about to spend $10 to find out whether AI images suit their work at all.
If you need images most days but they are not the product you sell, a ChatGPT subscription you already have covers it. Blog covers, social graphics, slide images, mockups. Adding a second $30 subscription for that is hard to justify.
If the images are the product, Midjourney. Concept art, editorial work, anything where a client is paying for how it looks rather than what it depicts. Its aesthetic tuning is a deliberate choice and it shows.
If you need readable text inside the image, test both this week rather than trusting any comparison including this one, for the reasons above.
If you need guaranteed commercial safety, neither of these is the automatic answer. Licensing is where AI image tools get people into trouble, and self-hosted open models are worse rather than better on this point despite feeling freer. I go through the licence terms model by model in Midjourney vs Stable Diffusion.
If you are new to Midjourney
The Discord workflow puts people off and the web app has largely fixed that, but it is still a dedicated tool with its own prompt language rather than something you can talk to. Atmospheric, descriptive prompts work better than structured ones, and parameters like aspect ratio and stylisation do real work once you understand them.
The step by step Midjourney guide covers setup. One caveat on it and on every other Midjourney tutorial you will find: most were written against v6 or v7, and specific parameter advice ages badly across major versions. Treat the workflow as durable and the exact numbers as suggestions.
The wider lesson, which is not about images
I have now written three posts this year about AI tools being switched off, and the pattern is consistent enough to plan around.
Voice tools shut down and took hosted audio with them. Now an image product retires and the vendor's advice is to download your files. Nothing here was a scandal. Products get replaced, and being replaced by something better is the normal case rather than the bad one.
What costs people real work is treating a hosted AI tool as permanent storage. If an output matters, it lives on your own disk, not in a conversation history you do not control. I wrote about the same failure mode in the ElevenLabs alternatives piece, where two of the tools most often recommended had already gone.
Verdict
For most people reading this, the honest answer is that this is no longer a $20 versus $30 decision. Image generation in ChatGPT is free to try and $8 to use lightly, and Midjourney starts at $10 with no free option. That changes who should even be comparing them.
Midjourney remains the better tool if you are selling how something looks. It has shipped three versions since I last tested it properly, and its aesthetic advantage comes from a product decision rather than a temporary lead, so I expect it holds.
OpenAI's side is now cheaper to start, bundled with something you may already pay for, and no longer called DALL-E. If your bookmark points at the DALL-E GPT, it stops working on August 30, and the images inside it are yours to save until then.
The thing I would not do is pick either one based on a comparison written against the previous generation of both. Including, until today, this one.
Verified on August 10, 2026, and the retirement status re-checked on August 31, 2026. The DALL-E retirement dates, the ChatGPT Images 2.0 capabilities, and the ChatGPT plan tiers come from OpenAI's own announcements and help pages. Midjourney's pricing and version history come from third party trackers, because midjourney.com and docs.midjourney.com both block automated access, so treat those figures as reported rather than verified.
