The other half of the audit. Setting aside what was lost — there are several things the 2026 models do that MidJourney v3/v4 in 2022-2023 could not do at all, or only barely.
Legible text in image.
The 2026 breakfast scene has a chalkboard reading "Today's Plan: Eat Pancakes / Be Happy / Repeat", an apron embroidered "Good Morning!", labeled bags of FLOUR and MILK, a "MAPLE SYRUP" bottle, and a sticky note reading "Have an awesome day!". Every word is correctly spelled. The 2022 original has no text and could not produce any. This is the single largest capability change between the two eras.
Likewise, the 2026 speedracer outputs put "RACCOON Speed Co." and "SPEEDRACCOON RACING CLUB" into the image as part of the design. In 2022 these would have been illegible scribbles. In 2026 they're typeset.
Multi-element composition that holds together.
The MJ raccoon-making-breakfast has a raccoon and a plate. The 2026 breakfast has a raccoon, a plate of pancakes with butter and berries, a frying pan with two eggs, a window with curtains, a chalkboard, a sticky note, a bowl of pancake batter and a whisk, a maple syrup bottle, a flour bag, a milk carton, a bowl of mixed berries, an egg carton, a houseplant, and a wood floor — and they're all spatially coherent. The MJ models could occasionally manage three or four discrete objects without errors; the 2026 models routinely manage a dozen.
Prompt adherence at the noun level.
Asked for a raccoon on the moon, MJ v3 gave a stylized raccoon on what reads as Mars. The 2026 ChatGPT version gives a raccoon in an astronaut helmet on a moon surface with the Earth in the background and a recognizable Saturn-V rocket. The Gemini version gives a smiling raccoon on the moon holding a flag with stars on it, plus a stylized cartoon Earth and a cartoon rocket. The 2026 models hit the literal nouns; MJ v3 hit a noun in the neighborhood.
Photoreal-without-asking.
The MJ raccoon-samurai is a print. The 2026 samurais are characters. You can read individual armor plates, fabric folds, sword sheen, dust kicked up by the leap. None of this is in the prompt — the prompt says only "cartoon racoon samurai." The new defaults treat "cartoon" as "highly polished animation key-frame" rather than as "watercolor illustration."
Speed and cost.
This pilot generated 70 images in roughly 90 minutes of wall-clock time, end-to-end, with zero paid API spend (both models were driven through their consumer chat UIs on existing accounts). The same volume of output through MidJourney in 2022 would have taken several hours of Discord babysitting and a paid subscription. The economic cost of image generation has effectively gone to zero for moderate volumes.
The full pilot — every collection, every comparison — is at the 2026 Recreation page. If you want to expand the pilot from 35 prompts to the full 1,662 in the archive, that's the next decision.