How we tested
All five models were tested between June 1 and June 18, 2026, on their current paid tiers (Midjourney Standard, ChatGPT Plus for GPT Image 1.5, Firefly Pro, Ideogram Plus, and direct API access for FLUX.2 [pro]). Scores reflect the model versions available in that window. Output quality is weighted heaviest, followed by prompt adherence; commercial-license clarity is weighted heavily for any pick aimed at brand or agency use.
Output Quality
Two reviewers independently scored the same set of 40 prompts (10 photorealistic portraits, 10 product shots, 10 illustrative scenes, 10 graphic designs) generated in each model, judging the best of four outputs per prompt against a fixed rubric covering composition, lighting, anatomy, and surface detail. We averaged the two reviewers' scores per prompt.
Prompt Adherence
We ran 20 deliberately compound prompts (each naming a subject, an action, a setting, a style, and at least three constrained details such as object color, count, or position) and counted how many of the named elements each model rendered correctly in the best of four outputs.
Text Rendering
We generated 15 graphics that required readable text in the image (a five-word slogan, a two-line poster, a product label, a sign with a numeral, a logo wordmark) and scored each output for spelling accuracy, letter integrity, and how legibly the text integrated with the composition.
Commercial-License Clarity
We read each vendor's terms of service and indemnification language and recorded whether the model is trained on disclosed licensed content, whether the paid tier carries written IP indemnification, and what the policy says about who owns the output and whether prompts may be used to train future models.
Cost per Usable Image
We priced one user on each tool's standard paid plan against the real ceiling of that plan (Fast GPU hours, credit allocation, or per-image API rate) and divided by the number of images a working creator generates in a month of professional use, including iteration, to record a realistic cost-per-keeper.
We ran every model through the same prompt battery, so the differences below sit in the products themselves, not the briefs. The per-criterion marks are above; the notes here cover where the ranking turned.
Why Midjourney still leads on the work we judged
Midjourney’s lead is narrower than it was two years ago, but it’s real on the dimension that decides this category for editorial and creative use.
Midjourney V7 introduced Omni Reference for precise character consistency and measurably improved photorealism
, and
the V8.1 update layered on top in late April 2026 brings faster generation, better prompt understanding, stronger small-detail retention, HD 2K image support, and Raw mode options
. The aesthetic is what subscribers come back for: on our portrait and concept-art prompts, Midjourney produced the strongest first draft more often than any other model.
The price is also more defensible than it looks.
The Standard plan costs $30/month, or $24/month annually, and it is the sweet spot for regular creators, you get 15 hours of fast GPU time plus unlimited images in Relax Mode
, and
Pro at $60/month adds Stealth Mode, which keeps generations private and hidden from Midjourney’s public gallery, important for client work
. The weaknesses are equally real: text rendering trails Ideogram badly, the Discord-first workflow is dated, and Midjourney offers no commercial indemnification.
When GPT Image 1.5 is the better all-rounder
OpenAI’s image work has consolidated.
DALL·E 2 and DALL·E 3 were removed from the API on May 12, 2026
, and the current model is GPT Image 1.5, which
leads quality benchmarks with an LM Arena Elo score of 1,264
. In our compound-prompt and text-rendering tests it was the most reliable model outside of Ideogram, and it remains the cheapest premium model to access.
ChatGPT Plus costs $20 per month and generates roughly 50 images every 3 hours through GPT Image 1.5, OpenAI’s native image model that replaced DALL-E 3
. For volume work, the API runs
at $0.04/image
for standard generations, with a budget GPT Image 1 Mini lane at $0.005 per image.
If you’re paying for ChatGPT already and need image generation, GPT Image 1.5 is effectively included. For most generalist work it’s the right first call.
When Firefly is the only defensible choice
Firefly is the model we recommend when a brand-legal review is part of the workflow.
Firefly is trained exclusively on licensed Adobe Stock content, openly licensed content, and public domain material, so Adobe is confident enough in its training data to back it legally
, and
part of what the enterprise tier buys is contractual indemnification: Adobe’s commitment to defend the customer against intellectual-property claims arising from Firefly-generated content
.
The enterprise add-on is where most large buyers land, at roughly $24 per user per month before volume discount
;
standalone Firefly plans run from Standard at $9.99/month to Pro at $19.99/month and Premium at $199.99/month
.
The trade-off is twofold. Firefly’s general aesthetic quality trails Midjourney and GPT Image 1.5 on portraits and editorial briefs, and the indemnity has real limits.
Adobe’s Generative AI User Guidelines prohibit prompts intended to produce content that infringes IP, defames real people, or imitates trademarked styles. Violate the guidelines, lose the indemnity.
Get the scope in writing for the seat type you’re buying.
When Ideogram is the right specialist
Ideogram 3.0 is the answer when readable text inside the image is part of the deliverable.
If your images need readable text, logos, social media graphics, posters, infographics, Ideogram is the only serious option. While Midjourney achieves roughly 30-40% text accuracy, Ideogram V3 hits 90-95%. That’s the difference between usable marketing material and gibberish.
The model’s typography pipeline treats letterforms as a first-class output, and on our 15 text-in-image prompts it was the only generator to ship every test legibly.
The pricing has tightened.
After Ideogram’s official plan docs in March 2026, the current ladder is more nuanced than older reviews suggest. The Basic plan still exists only as a legacy subscription for existing users; new buyers now choose between Free, Plus, Pro, and Team. For new users, the entry point is the Free plan with 10 slow credits per week. The first paid plan that feels commercially practical is Plus at $20/month or $180/year. It includes 1,000 priority credits per month, unlimited slow credits, private generation, uploads, the editor, and unlimited canvases. Pro jumps to $60/month or $504/year with 3,500 priority credits per month.
It’s a second tool in most production stacks, not a first, but it’s the second tool you actually need.
Why FLUX.2 is the developer’s pick, not the creator’s
FLUX.2 [pro] is the strongest model in our test for raw photorealism, and on aggregate quality benchmarks it sits within noise of GPT Image 1.5.
A pricing calculator on BFL’s site indicates that FLUX.2 [Pro] is billed at roughly $0.03 per megapixel of combined input and output. A standard 1024×1024 (1 MP) generation costs $0.030, and higher resolutions scale proportionally.
FLUX.2 generates high-quality images while maintaining character and style consistency across multiple reference images, following structured prompts, reading and writing complex text, adhering to brand guidelines, and reliably handling lighting, layouts, and logos. It can edit images at up to 4 megapixels while preserving detail and coherence.
The reason it lands fifth, recommended but at the back of the recommended field, is workflow. There’s no polished first-party web app for non-developers, no consumer subscription, and provider availability is still uneven.
as of March 2026, FLUX.2 [pro] is only available through BFL’s official API and Replicate. Major platforms like fal.ai have not yet launched FLUX.2 [pro] support, leaving developers with few options for competitive pricing.
For a product team building generation into an application, FLUX.2 is excellent. For a working creator who wants to open a browser and make an image, the rest of this list serves you better.
What did not make the cut
We tested but did not recommend Stable Diffusion XL and the self-hosted FLUX.2 [dev] open-weight checkpoint. Both are credible for technically capable teams.
FLUX.2 [dev] is a 32B open-weight model, the most powerful open-weight image generation and editing model available today, combining text-to-image synthesis and image editing with multiple input images in a single checkpoint, with weights available on Hugging Face
But neither delivers a serious managed experience for a working creator, and frontier quality on SDXL trails the hosted leaders by a meaningful margin in 2026. They’re infrastructure picks, not products we can recommend to a reader who wants to open a tool and make an image today.
