This pilot puts each prompt through three engines: MidJourney v3/v4 from 2022-2023 (the original generation), Google Gemini in 2026, and ChatGPT in 2026 (the GPT-image-1 model under the hood). Most observations about "how AI art changed" turn out to be unprovable hand-waves once you sit with the actual side-by-sides. Here are four that survived.
The 2022 image is small. The 2026 images are loud.
The MJ samurai is a tiny figure inside a poster — most of the frame is the city. The ChatGPT samurai fills the frame. The Gemini samurai fills the frame. In every triplet I checked, the 2022 image gives the subject roughly 20-40% of the canvas; the 2026 images give it 60-90%. The new models read every prompt as a portrait of the noun.
2022 understood subtraction. 2026 piles on.
"Cartoon racoon making breakfast" → MJ delivered a quiet, painterly raccoon holding one plate. ChatGPT delivered a kitchen with eggs, pancakes, butter, maple syrup, three labeled containers, an apron, a chalkboard to-do list, a sticky note, a window plant, and a wood floor. Same six-word prompt; two different ideas of what "complete" means.
You see the same impulse on the speedracer-logo prompt:
MJ's reading: "logo." 2026's reading: "vintage racing-club merchandise badge with copy." Neither is wrong — Mad Bitcoins did keep returning to the prompt 20 times, presumably because he wanted something closer to MJ's reading — but the 2026 default has moved.
The new models read text-in-image as part of the brief.
Looking across the four raccoon triplets above: every 2026 image has legible English text. The Gemini moon-raccoon holds an "I ♥ THE MOON!" flag. The ChatGPT samurai-raccoon stands in front of a building with what reads as legible Japanese signage. The breakfast kitchen has FLOUR, MILK, Maple Syrup labels and a chalkboard. The speedracer logo has "RACCOON Speed Co." and a turbo gauge labeled "TURBO." The MJ originals do not have any of this — when they have characters, they're ornamental scribbles.
This is the cleanest "2026 does X that 2023 couldn't" claim — and it changes what these prompts mean. A 2022 prompt assumed text was off the table. A 2026 prompt about, say, "Capitol building with huge bitcoin logo" now has to specify whether you want the logo or not, because the model will paint it on three buildings.
The MJ mood is gone.
The original MJ raccoons are somber. The on-the-moon raccoon looks sad. The making-breakfast raccoon looks downcast. The samurai raccoon scowls behind a flat black mask. The 2026 raccoons smile. They wave. They give thumbs-up. They wink. This isn't a technical change — it's a tonal default. ChatGPT in particular routes nearly every cartoon-animal prompt to a "friendly mascot" register, which is fine for a pancake-flipping breakfast scene and odd for a samurai mid-strike.
For Mad Bitcoins' archive specifically, that tonal shift may matter more than the resolution bump. The originals had a wistfulness that read as art. The recreations have a polish that reads as product. Neither is better. They're just different jobs.