Comparisons & Benchmarks

Comparisons & Benchmarks

Midjourney, GPT Image 2, and Flux 2

If you came here searching for “Midjourney V7 vs DALL-E 4 vs Stable Diffusion 4,” stop before the rest of this wastes your time: two of those three products don’t exist under those names anymore, and the third has already moved past the version most articles are still testing. That’s not a nitpick. It changes which tool wins — and it changed again three weeks after this piece was first published, which is the whole reason it’s worth reading a dated comparison instead of a static one.

What’s actually shipping, as of today

Here’s the current lineup, not the one still floating around in search results from earlier this year.

Two release dates are worth being precise about, since a prior version of this piece wasn’t. Midjourney’s V8 line moved fast: V8 alpha shipped March 17, 2026; V8.1 alpha followed April 14, 2026, and became the site-wide default June 10; V8.2 then took over as default on July 24, after a preview window that opened June 25 ESTABLISHED (source). Unlike V8.1, which was primarily a speed release, Midjourney describes V8.2 as an aesthetics-and-personalization update — bolder, more contrast, fewer low-quality outputs — with no claimed change to style-reference consistency across a series ESTABLISHED.

Where each one actually wins

Midjourney V8.2: still the one people can identify blind

Midjourney’s aesthetic — deliberate color grading, strong compositional intent, a look that reads as designed rather than photographed — is still recognizable at a glance. V8.2 pushes that further toward “bold and sophisticated” by the company’s own description, at the cost of some of V8.1’s more restrained, V7-inspired output ESTABLISHED. Where it still lags: on-image text remains limited, there’s still no public API and no way to fine-tune it on your own dataset, and Midjourney has never disclosed model architecture or entered the blind-vote arenas the rest of this piece leans on for evidence (more on that below). If the brief is “make this look expensive,” it’s still a strong default. If the brief needs a client’s exact logo rendered correctly, or an automated pipeline, it isn’t.

GPT Image 2: the text-and-precision leader

This is the real headline of 2026’s image-gen cycle, and it’s underreported because “DALL-E” is what people still type into Google. GPT Image 2 added a “thinking mode” reasoning step that plans composition and checks constraints before rendering ESTABLISHED (source). OpenAI’s own materials claim roughly 99% character-level text accuracy across Latin, CJK, Arabic, and Hebrew scripts VENDOR-CLAIMED — that specific figure comes from OpenAI-aligned marketing rather than an independently audited benchmark, so treat it as directional rather than exact. The more defensible, independently-sourced version of the same claim: on the Artificial Analysis Image Arena, GPT Image 2 debuted at #1 across all seven text-to-image sub-categories, with its largest single-category jump in text rendering specifically ESTABLISHED (source). Both point the same direction; only one of them is a number you should quote in a client deck.

The catch is cost predictability. Pricing is token-based, not a flat per-image fee: OpenAI lists $5 per million input text tokens, $10 per million output text tokens, $8 per million input image tokens, and $30 per million output image tokens ESTABLISHED (source). In practice, third-party trackers put that at roughly $0.02–$0.21 per image depending on resolution and reasoning depth — you can’t quote a single number the way you can for a flat-rate API, and that makes budgeting for high-volume production genuinely harder.

Flux 2 and SD 3.5: the self-hosted, no-per-image-fee lane

If your actual constraint is “we generate too many images to pay per image,” this is the category that matters, and it’s the one the “vs Stable Diffusion 4” framing gets most wrong. Stability AI hasn’t shipped a successor to SD 3.5 since October 2024 ESTABLISHED (source) — the open-weight energy in this space moved to Black Forest Labs, founded by former Stability AI researchers, whose Flux 2 family launched in November 2025.

Flux 2’s pricing is tiered by megapixel, not a flat rate, and it’s worth stating precisely because a version of this claim was flattened in an earlier draft. Per Black Forest Labs’ own documentation: Klein 4B charges $0.014 for the first megapixel and $0.001 for each additional megapixel; Pro charges $0.03 for the first megapixel and $0.015 for each additional one, with input reference images billed separately at $0.015/MP ESTABLISHED (source). A standard 1024×1024 (roughly 1MP) generation on Pro lands around $0.03; on Klein 4B, around $0.014. Multi-megapixel or multi-reference jobs cost more than either headline number suggests — use BFL’s own pricing calculator before budgeting a batch job, not a single quoted rate.

Klein’s 4B tier is Apache-licensed and runs on consumer GPUs with as little as 13GB VRAM; Dev is open-weight but non-commercial by default unless you buy BFL’s self-hosted commercial license; Pro and above are hosted commercial API tiers ESTABLISHED. SD 3.5 remains free to self-host and free commercially for organizations under $1M in annual revenue under the Stability AI Community License ESTABLISHED (source). Neither will out-polish Midjourney on stylized work or out-render GPT Image 2 on dense text, but if the job is high-volume product photography or brand imagery where you need to own the weights, this is where the value actually sits.

What the blind-vote leaderboards say — and where they disagree

A prior version of this piece made a photorealism claim about Flux 2 without citing the blind-vote data that actually exists. Here it is — along with the honest complication: the major leaderboards don’t agree with each other on absolute numbers, and one major vendor isn’t on them at all.

Three blind-vote platforms track these models by pitting two anonymous outputs against each other and aggregating human votes into an Elo-style score: the Artificial Analysis Text-to-Image Arena, the LMArena (arena.ai) text-to-image board, and the aggregated llm-stats.com board. All three currently rank GPT Image 2 first among tracked models ESTABLISHED. Where they diverge is the score itself — Artificial Analysis and LMArena reported GPT Image 2’s launch-week Elo above 1,500 in April 2026 on a scale where Flux 2 Pro sat in the low-to-mid 1,100s and SD 3.5 Large around 940 ESTABLISHED (source), while llm-stats.com’s differently-normalized scale showed GPT Image 2 at 485 as of late August 2026 with GPT Image 1.5 second at 290 ESTABLISHED (source). Elo scales aren’t standardized across platforms, so don’t compare raw numbers between boards — compare each platform’s internal ranking order instead.

The bigger gap: Midjourney does not participate in any of these three arenas. It has no public API, and Elo-style blind testing depends on programmatic access to generate paired outputs at scale ESTABLISHED (source). That means every “Midjourney beats X” or “X beats Midjourney” claim you read — including comparisons on this page — is a qualitative judgment from reviewers, not a blind-vote result. It’s a real limitation of the evidence, not a reason to dismiss Midjourney’s output quality, which independent reviewers consistently rate as excellent on aesthetic grounds. Just don’t let a specific Elo number get attached to it; none exists.

Speed: the trade-off nobody puts in the headline

Quality comparisons dominate this space; latency mostly gets left out, and it shouldn’t — it’s often the deciding factor for batch and production work.

This is why “which tool is best” and “which tool is best for a 5,000-image batch job” are different questions. GPT Image 2’s reasoning step is exactly what makes it reliable at rendering text and following complex layout instructions — and exactly what makes it a worse fit for high-volume, low-latency generation than Flux 2’s Klein tier, which was explicitly built for that use case ESTABLISHED (source). If you’re building a pipeline, benchmark p99 under your expected peak load, not the vendor’s best-case single-request number — API rate-limit tiers mean queue wait, not model speed, often dominates real-world latency at scale ESTABLISHED (source).

The unpopular take

The “vs” framing that drives most of these articles — including two earlier drafts of this one — assumes you’re picking one tool forever, and that the picture holds still long enough to make a permanent call. It doesn’t: Midjourney changed its default model twice in the six weeks after this piece’s first publication date. In practice, the teams getting the best output are running two tools at once: Midjourney or Flux 2 for the visual direction, GPT Image 2 when the deliverable needs legible on-image text or precise instruction-following. Loyalty to one image model in 2026 is mostly a budgeting decision, not a quality one, and treating it as a permanent choice is how you end up defending a tool that’s fallen a version behind without noticing.

When not to use each one

Skip Midjourney if the deliverable needs accurate on-image text, a specific brand identity baked in, or programmatic/API access — you’ll fight the model more than you’ll direct it. Skip GPT Image 2 for high-volume, low-budget batch generation unless you’ve actually modeled the token cost first, and unless your workflow can tolerate thinking-mode latency; the flat-rate, low-latency competitors are meaningfully cheaper and faster at scale. Skip Flux 2 Dev and SD 3.5 if you don’t have someone who can manage local GPU inference or a hosted endpoint — the cost savings evaporate the moment you’re paying an engineer’s hourly rate to keep it running.

Quick reference: mythe & fact

FAQ

Elo score

A relative-ranking system borrowed from chess ratings; used here to aggregate blind human votes between pairs of AI-generated images. Not comparable across different platforms’ scales.

Thinking mode

GPT Image 2’s optional reasoning step that plans image composition and checks prompt constraints before rendering, trading latency for accuracy.

Megapixel-based pricing

Black Forest Labs’ billing model, where cost scales with output (and sometimes input reference) resolution rather than a flat per-image fee.

The tool didn’t change your standards. The version number did — twice, in the six weeks since this piece first went up — and most of the internet still hasn’t caught up.