DALL-E is dead. OpenAI retired the DALL-E brand and replaced it with GPT Image 1.5. Midjourney just launched a V8 alpha while V7 remains the default. And Flux 2 Pro keeps climbing the LM Arena leaderboard, with an open-weight version you can run at home.
We put all three through the same set of prompts. Here’s what actually happened.
The Contenders
GPT Image 1.5 — OpenAI’s native multimodal image generator, built directly into the same model that handles your text. Rolled out globally December 2025, available to all ChatGPT tiers. Unlike DALL-E, which used a separate diffusion model, GPT Image 1.5 generates images inside the same neural network processing your prompt. That architectural change means it understands what you want with less hand-holding.
Midjourney V7 — The current default model, known for producing images that look like they were shot by a professional photographer or crafted by a skilled digital artist. V7 brought improved hand and body coherence, smarter prompt interpretation for longer descriptions, and a personalization system that learns your aesthetic preferences over time.
Flux 2 Pro — Black Forest Labs’ flagship, a 32-billion-parameter model laser-focused on camera-accurate visual characteristics. Depth of field, lens distortion, chromatic aberration, film grain — Flux doesn’t simulate these effects, it reproduces them with optical precision. Generates images in 3–5 seconds.
Test 1: Photorealism
Prompt: “A middle-aged woman sitting at a café terrace in morning light, reading a newspaper, shallow depth of field, shot on 85mm lens”
GPT Image 1.5 nailed the prompt literally. The newspaper was readable. The morning light was warm but not overdone. Shallow depth of field worked correctly. It looked like a stock photo — technically perfect, emotionally flat.
Midjourney V7 took the same prompt and added atmosphere. The light had golden hour quality. The woman’s expression had a story behind it. The background blur felt natural. Less technically precise than GPT Image, but the kind of photo you’d actually stop scrolling for.
Flux 2 Pro split the difference. It captured the 85mm lens distortion accurately — you could almost identify the lens model from the bokeh pattern. Skin texture and fabric detail were the best of the three. The overall look was closest to a photo from a high-end DSLR.
Winner: Flux 2 Pro for pure photorealism. Midjourney V7 if you want the photo to feel like something, not just look like something.
Test 2: Text Rendering
Prompt: “A vintage movie poster for a film called ‘THE LAST ALGORITHM’ with a tagline ‘Some equations can’t be solved’ in smaller text below”
This is where GPT Image 1.5 dominated. Title text: perfect. Tagline: perfect. The poster design itself was coherent, with proper hierarchy between the title and tagline. OpenAI’s native multimodal approach gives it a significant edge here — the same model understanding both the image composition and the text content.
Midjourney V7 got the title right but garbled the tagline, rendering it as “Some equtions can’t be sovled.” Classic Midjourney text problems, though they’ve improved since V6.
Flux 2 Pro handled the title correctly and got 80% of the tagline right, with one letter swap. Considerably better than Midjourney but still not GPT Image territory.
Winner: GPT Image 1.5, by a wide margin. If text rendering matters, this is the only serious choice among these three.
Test 3: Complex Scene Composition
Prompt: “An isometric view of a tiny cyberpunk ramen shop at night, neon signs in Japanese, steaming bowls on the counter, a robot chef, rain on the windows, detailed interior”
All three generators produced interesting results here, but in very different ways.
GPT Image 1.5 followed every element of the prompt. Robot chef — present. Neon signs — there, and mostly legible in Japanese. Rain on windows — visible. But the composition felt like a checklist. Everything was there, nothing surprised you.
Midjourney V7 took creative liberties. The neon glow bled realistically through the rain. The isometric perspective was slightly off but in a way that made it feel like concept art, not a technical diagram. The robot chef had personality. The ramen shop told a story about the world it existed in.
Flux 2 Pro produced the most technically consistent isometric view. Clean lines, accurate perspective, good detail in the interior. The rain effects were photorealistic. But compared to Midjourney’s version, it felt like an architectural rendering rather than concept art.
Winner: Midjourney V7. Complex creative scenes are still where Midjourney’s artistic DNA shows most clearly.
Test 4: Speed and Iteration
This isn’t about a single image — it’s about the workflow of refining an idea.
GPT Image 1.5 generates in about 10–15 seconds per image. The conversational interface means you can say “make the sky more dramatic” and it understands context from previous generations. That iterative loop is powerful.
Midjourney V7 takes 30–60 seconds per image in standard mode, though Draft Mode cuts that to about 6 seconds at half the cost. The V8 alpha reportedly does 5x faster generation. The upscale and variation workflow is well-designed once you know it.
Flux 2 Pro generates in 3–5 seconds via API. The fastest of the three for raw generation, though it lacks the built-in iterative refinement that GPT Image and Midjourney offer.
Winner: GPT Image 1.5 for iterative workflow. Flux 2 Pro for raw speed.
The Pricing Reality
| Generator | Cost | Notes |
|---|---|---|
| GPT Image 1.5 | $0.04–$0.12/image (API); included in ChatGPT Plus ($20/mo) | Token-based billing, cost varies with complexity |
| Midjourney V7 | ~$0.033/image in Fast mode; unlimited in Relax | Plans: $10/mo Basic, $30/mo Standard, $60/mo Pro |
| Flux 2 Pro | $0.04–$0.055/image (API) | Or run Flux 2 Dev/Klein locally for free |
Midjourney’s Relax mode makes it effectively unlimited for $30/month if you can tolerate the queue. GPT Image 1.5 is included in ChatGPT Plus but hits usage caps. Flux 2 Pro costs per image via API — but you can self-host the open-weight alternatives for nothing.
The Privacy Angle Nobody Talks About
Here’s where these three generators diverge sharply.
GPT Image 1.5: Your prompts and generated images flow through OpenAI’s servers. With ChatGPT, your interactions can be used for model training unless you opt out. Every image you generate feeds OpenAI’s data pipeline.
Midjourney: Operates on a training data model built from publicly crawled web images. Your generated images default to public unless you’re on a Pro or Mega plan with Stealth Mode. Their privacy policy covers data collected “through the process of training Midjourney machine learning algorithms.” The February 2026 joint statement from 61 data protection authorities specifically flagged AI image generators capable of creating realistic depictions of identifiable individuals.
Flux 2: This is the privacy win. Flux 2 Dev and Flux 2 Klein are open-weight models you can run entirely on your own hardware. The Klein 4B model needs roughly 13GB of VRAM — an RTX 3090 or 4070 handles it. Your prompts never leave your machine. Your images stay on your disk. The 4B parameter model is Apache 2.0 licensed — fully open for commercial use, modification, and redistribution. Ollama even added experimental Flux 2 Klein support for macOS.
If you’re generating images of products, prototypes, unreleased designs, or anything proprietary, running Flux locally means no third party ever sees your work.
The Leaderboard Says…
The LM Arena text-to-image leaderboard — the closest thing to an objective ranking, based on human preference votes — puts GPT Image 1.5 at the top with an Elo score of 1264. Flux 2 Pro sits right behind it at 1265 (statistically tied). Midjourney V7 scores lower on the leaderboard, but leaderboard rankings favor prompt adherence and photorealism over artistic quality — exactly the areas where Midjourney deliberately trades precision for atmosphere.
The 2026 VibeDex benchmark tells a different story for text-in-image work: Ideogram 3.0, which once led the field in text rendering, has dropped to 10th place as GPT Image 1.5 and others have caught up and passed it.
What About Midjourney V8?
The V8 alpha launched March 17, 2026 as a preview on alpha.midjourney.com. It brings 5x faster generation (under 10 seconds), native 2K resolution via —hd mode, improved text rendering, and a new Style Creator for extracting reusable visual styles from reference images. It’s built on an entirely new GPU-native codebase.
Early reports suggest V8 narrows the gap with GPT Image on text rendering and with Flux on photorealism, while maintaining Midjourney’s signature aesthetic edge. It’s not the default yet — V7 still is — but it signals where the platform is heading.
Our Verdict
There’s no single best AI image generator in April 2026. There are specialists:
Choose GPT Image 1.5 if you need reliable text rendering, want the easiest iterative workflow, or already live inside ChatGPT. It understands complex prompts better than anything else and the conversational refinement loop is unmatched.
Choose Midjourney V7 if you want images with artistic impact. Concept art, mood boards, creative direction, anything where the image needs to evoke a feeling rather than reproduce a specification. The aesthetic quality remains a step above everyone else.
Choose Flux 2 Pro if you need photorealism, care about privacy, or want to run things locally. Flux 2 Pro via API matches GPT Image on technical quality. Flux 2 Dev and Klein let you self-host on consumer hardware with zero data leaving your network. For privacy-conscious work with proprietary content, nothing else comes close.
The dark horse option: Run Flux 2 Klein locally for daily work (free, private, fast), use Midjourney for hero art that needs to stand out, and hit GPT Image 1.5 when you need text baked into the image. Three tools, three strengths, zero overlap.
What You Can Do
-
Try all three — GPT Image 1.5 is in ChatGPT (free tier gets limited access). Midjourney’s Basic plan is $10/month. Flux 2 Pro offers API access through multiple providers.
-
Run Flux locally — If you have 13GB+ VRAM, grab Flux 2 Klein from Hugging Face and run it through ComfyUI or Ollama. No cloud, no costs, no data sharing.
-
Audit your current tool — Check what happens to images you generate. Does the platform use them for training? Are they public by default? Can you delete them? These questions matter more than they did a year ago — 61 data protection authorities said so in February.
-
Watch the V8 alpha — If you’re a Midjourney user, it’s worth testing V8 on alpha.midjourney.com. The speed and text rendering improvements could change your workflow.