Peter Yang compares Fable 5.2 with Opus 5.5 in five creator-run tasks. His aesthetic preferences and small repeated writing and pushback tests are useful observations, not an independently reproduced ranking.

Visual work: three different tasks

  • Kyoto promo video: both models write code for a 20-second Fushimi Inari travel reel, with an initial result and one self-critique iteration. Yang prefers Fable’s gate sequence and finish; he says rendering occupied much of the time.
  • Kart racer: both build a single-track browser game using Three.js and self-review once. Yang reports 34 minutes for Opus and 62 for Fable, prefers Fable’s animation, then gives Fable three additional polish rounds. The final reel is not the original two-iteration comparison.
  • Website redesign: the bold-design prompt produces visually distinctive pages; Yang likes Fable’s theater theme but spots a misplaced call to action. A separate conversion-focused prompt brings the designs closer. He prefers Fable slightly, yet still favors his existing site; no conversion experiment is shown.

Writing and disagreeing with the user

  • Each model writes five 600-word essays from the same prompt without Yang’s custom style instructions. His phrase-counting check finds few expressions he calls “AI slop” in either model; in a personal blind comparison he picks Fable in four of five pairs and notices Opus repeats the same legal-research story.
  • Both models identify a missing course-activation goal in his business plan. In repeated challenges, both resist plain disagreement and claimed social proof; Yang says Opus more often holds its ground under emotional pressure and a false claim that usage numbers improved.
  • His proposed AGENTS.md/CLAUDE.md instruction asks for candid, plain writing. He says it helps both models, but the video does not report a separately controlled before/after evaluation of that instruction.

“I actually prefer my existing version better than what both models came up with.” — Peter Yang, on the website task