Midjourney

Where it sits

Midjourney is the image generator everything else in this category gets measured against, and in 2026 it is also one of the strangest companies in software. Founded by David Holz in 2021, launched to the public on July 12, 2022, headquartered in San Francisco, and never funded by a single outside investor — Holz has repeatedly declined venture money, and estimates compiled by analysts put 2025 revenue around US$500 million with a team of roughly 100–160 people. Whether those third-party revenue figures are exact is debatable; the direction is not. Midjourney sells subscriptions for GPU time, has been profitable since about a month after launch, and answers to no board. In August 2025 Meta licensed its image and video technology for use across Facebook, Instagram and WhatsApp — a licensing deal, Holz stated, with no equity involved and no change in Midjourney's independence.

That independence shows up in the product. Midjourney ships when its models are ready, on its own cadence, with an aesthetic point of view. The default model is V8.1 as of June 2026 (V8.2 landed in July 2026), and the pace has been relentless: V7 became default in April 2026, V8.1 followed in June, a V8.2 edit model entered testing in August 2026, and the company is currently trialing what it calls Thinking Mode.

What it actually does

Midjourney generates images from text prompts and, increasingly, images from images. The current feature set includes text-to-image, image editing with natural-language instructions, up to four image references at once through the Edit model (which is replacing the earlier omni-reference system, --oref, with its 0–1000 strength slider), inpainting and outpainting, style references (sref), moodboards, personalization that learns your taste from ratings, and video generation (text- and image-to-video, SD on Basic and HD-capable on Standard and above, with concurrent video jobs scaling 1/3/6/12 up the plan ladder). The web app at midjourney.com is now the primary interface; Discord still works and retains the parameter syntax that power users build workflows around. Draft mode burns less GPU time for quick iterations, HD mode renders at what Midjourney describes as twice the size and 4x the resolution of V7 output, and Midjourney claims V8.1 renders an SD image in about 4 seconds and an HD image in about 12 — vendor speed claims, but in my experience the queue is fast in practice.

Pricing, at time of writing, from docs.midjourney.com/docs/plans: Basic US$10/month (US$8/month billed annually at US$96), Standard US$30 (US$24 annually at US$288), Pro US$60 (US$48 at US$576), Mega US$120 (US$96 at US$1,152). What you are buying is Fast GPU time — 3.3, 15, 30 and 60 hours per month respectively — with unlimited Relax Mode image generation on Standard and above and unlimited SD video Relax on Pro and above. Extra Fast time is US$4/hour on every plan. Stealth Mode, which keeps your generations out of the public gallery, exists only on Pro and Mega. Concurrency climbs from 3 concurrent Fast image jobs to 12 on Pro/Mega. Companies grossing over US$1 million per year must be on Pro or Mega for commercial use. Two smaller mechanics matter day to day. Rating images earns free Fast GPU time — the top 2,000 raters each day get an extra hour, and awarded time requires an active subscription and expires after 30 days — which keeps the community gallery feeding the personalization system. And repeat/permutation batching scales with tier (up to 40 jobs on Pro/Mega), which is how I run style sweeps across a parameter range without babysitting the queue.

My testing notes

Two concrete observations from my recent testing. First, the aesthetic gap is real but narrower than it was. On a controlled set of prompts — a product shot on colored paper, an editorial portrait, a stylized poster — V8.1's default output had the best lighting coherence and composition of any generator I ran this cycle, and the text rendering, historically Midjourney's weak spot, is now usable for short headlines (it still garbles long strings where Ideogram does not). Second, the metering model changes how you work. On Standard, the moment Fast hours ran low I shifted to Relax Mode and the workflow simply continued slower — that "unlimited slow lane" is the single most valuable feature on the plan ladder, and it is why Standard is the tier almost everyone should buy.

Personalization is the feature I underestimated. After a week of rating images and building a small moodboard, my default generations started tracking my own taste — warmer palettes, tighter crops — without any prompt changes. For a solo designer this compounds into a house style you own; for a team it means every member's output subtly diverges, so studios should pin sref codes rather than rely on personal defaults. That is a workflow decision Midjourney does not make for you.

I also tested the reference pipeline. Feeding the same character through the new edit model with up to four reference images preserved identity across outfit and background changes better than the old omni-reference flag, though the docs themselves admit V7 omni-reference remains available while the V8 version finishes training — the feature surface is genuinely mid-migration and some paths still route through older machinery.

Where it falls short

There is no free tier and no free trial; Midjourney removed free access in March 2023 and the only entry point is US$10 for a month. There is no public API — Midjourney has never offered one, so if you need programmatic generation you are looking at Flux, Stability or OpenAI instead (third-party resellers exist and are against the terms). Privacy costs extra: everything you generate is public in the community gallery unless you pay for Pro. The credit of "best aesthetics" comes with an opinionated style — Midjourney wants to make images look a certain way, and fighting its house style with heavy negative prompting is less effective than switching tools. Long text in images remains unreliable. Enterprise buyers get no SLA, no SSO, and a support model built around Discord and community. And the company is a bet on a small team: roughly 100–160 people, no enterprise sales organization, and a roadmap that just added a hardware project (Holz has said the "Orb" program may require outside capital for the first time). The public-gallery default is the quiet compliance problem: on Basic, Standard and anything below Pro, every generation is visible in the community feed, which is a non-starter for unreleased product work regardless of what the license allows.

Verdict

Rated 4.5 on the ToolSpotter rubric: core functionality depth is at the top of the category for image generation and now credible for video and editing (the 40%); value is good at US$30 Standard given unlimited Relax generations but weakened by the absence of any free tier (the 20%); usability is excellent on the web app and idiosyncratic in Discord (the 20%); the community, style ecosystem and sref/moodboard tooling are unmatched (the 10%); and the company's self-funded profitability and Meta licensing deal are strong reliability signals tempered by small-team concentration (the 10%). If image quality is the job, this is still the one to beat. If you need an API, privacy on a budget, or a free way in, look at the rest of this list.

Sources

• https://docs.midjourney.com/docs/plans
• https://www.midjourney.com/updates/announcements
• https://updates.midjourney.com/edit-model-for-v8/
• https://alpha.midjourney.com/updates
• https://aiwiki.ai/wiki/midjourney
• https://axis-intelligence.com/midjourney-statistics
• https://thefoundernation.com/why-midjourney-stayed-profitable-without-vc-funding
• https://www.toolcolumn.com/pricing/midjourney-pricing
• https://aipricecompare.org/apps/midjourney.html
• https://kenerateai.com/blogs/is-midjourney-free
• https://dev.to/stimlau/midjourney-pricing-explained-standard-pro-and-mega-in-2026-33n8