Peter Yang tests four models across five creative and knowledge-work tasks in this 19-minute comparison. The rankings below are his judgments of the demonstrated outputs, not independently verified performance claims.

Code-built scenes and games

  • Himeji Castle: Opus leads on architecture and cinematic polish; Gemini follows, then Astra and Muse.
  • Star Fox × StarCraft: Opus produces the most polished shooter; Gemini adds more features but crowds the experience.
  • Astra’s shooter has atmosphere but fewer enemy types and simpler backgrounds; Muse’s version looks substantially weaker.
  • The published game recipe includes a follow-up prompt, so “one-shot” should not be read as a fully controlled protocol.

Storytelling is not graphics quality

  • Opus delivers the strongest animation, but Yang is less convinced by its story about someone waiting for a spouse.
  • Gemini’s short is artistically strong; Astra’s plant-watering robot introduces an unintentionally comic visual.
  • Muse wins his emotional-story test: an old man’s gardening ends with a flower placed at a presumed spouse’s grave.

Website design: restraint beats more features

  • Astra wins: a clean layout, clear email signup, and generated illustrations support the page’s purpose.
  • Opus adds a playful peel-back-corner interaction but leaves the rest comparatively plain.
  • Gemini supplies a design system yet overfills the page; Muse produces a text-heavy layout Yang finds dull.
  • Image-generation access differs between the demonstrated workflows, which complicates model-only comparisons.

Writing and research

  • For interview-to-newsletter writing, Gemini wins an Opus subagent’s comparison against Yang’s human-edited final copy.
  • The judge favors Gemini’s personal voice and examples; Opus is insightful but less personal, and Astra is accurate but less distinctive.
  • Muse is not tested on writing; a separate judging agent is still not a blind, repeatable human evaluation.
  • Yang also prefers Gemini’s detailed medical-research answers—an anecdotal preference, not evidence of clinical accuracy.

What to take away

  • Pick models using representative tasks and explicit criteria: visual polish, emotional effect, usability, and voice are different goals.
  • Reproduce the prompts and inspect the outputs rather than treating one selected demo as a universal ranking.
  • The video does not establish repeated-run reliability, equal tool budgets, latency, or total cost.

“I think it really comes down to the use case.” — Peter Yang, 17:47