The single most demonstrable change between MidJourney v3 and the 2026 image generation is text. MidJourney v3 could not render English. It produced texture that looked like writing if you didn't read it — letterforms that were the shape of letters and the spacing of words but were, on inspection, illegible noise. The 2026 models read, write, and spell.
Every single 2026 image in this pilot contains at least one piece of legible English. Most contain several. The text appears on chalkboards, signage, t-shirts, banner graphics, building facades, video-game UI, comic-book panels, and protest signs. The text is in the prompt sometimes — but mostly it is the model's own contribution, added because the model has learned that an image with words feels finished.
The chalkboard
The canonical example, again, is breakfast. The 2022 MJ raccoon stands against a beige ground holding an unreadable plate. The 2026 ChatGPT raccoon has built him an entire labeled kitchen: FLOUR, MILK, an apron embroidered Good Morning!, and the chalkboard:
None of that text was in the prompt. The prompt was four words: cartoon racoon making breakfast. The model wrote Today's Plan: Eat Pancakes / Be Happy / Repeat on its own initiative, then wrote it cleanly enough to be readable. Three years ago that signage would have been illegible jitter.
The logo, as text
The 2022 prompt "a racoon version of speedracer in a circle for a logo" was Mad Bitcoins asking for a literal logo — clean geometric raccoon-in-circle, three colors, like a flag patch. MJ v3 gave that back. The 2026 models interpret "logo" as "branded sticker with display text," and the display text works:
ChatGPT delivers RACCOON Speed Co., a turbo gauge, a number 7. Gemini delivers SPEEDRACCOON RACING CLUB. Both invented brand names. Both spelled them correctly. The kerning is good. The badge would print.
UI and signage
The text-handling shows up most clearly when the prompt asks for a screen or sign. "16 bit mad hatter video game" in 2022 produced a pixelated illustration with garbled text. In 2026, both models produce legible UI: the title is the title, the score is the score, the dialogue box says something that parses.
The capitol-with-bitcoin-logo prompt makes the typography point overtly. The 2022 MJ version of that prompt simply could not render the bitcoin logo as text. ChatGPT 2026 puts a sharp, legible Bitcoin symbol on the building, larger than the Capitol dome.
What the model added on its own
The legibility goes both ways. Sometimes the prompt asks for text and the model delivers. More often, the model decides on its own that a scene needs words — a storefront sign, a t-shirt, a chalkboard, a wall mural — and writes them in. Three years ago the model would have hallucinated typography-shaped noise. Today it writes a full sentence and signs the kitchen with it.
This is what most readers will see first when they flip through the pilot: a 2022 archive that is texturally interesting but functionally illiterate, and a 2026 set where every image has a caption baked into it. The image learned to read. That is the most measurable thing that changed.