Peter Yang compares GPT-6.2 Astra, Claude Fable 5.2 and Gemini 4 Carbon across five practical tasks in this 23-minute video.

1. Render a rainy Tokyo travel video

  • Each model receives the same brief: a 20-second Shibuya Crossing video, built in code with music and a descending camera move.
  • Yang reports 11 minutes for Astra, almost an hour for Fable and 73 minutes for Carbon.
  • He prefers Fable’s camera work and street-level scene; crowds and traffic remain weak across all three.

2. Build a playable monster game

  • The brief combines Three.js, a browser graphics library, with Blender models and destructible buildings.
  • Yang reports 13 minutes for Astra, 47 for Fable and 49 for Carbon; each gets an initial build and one feedback round.
  • He ranks Fable first for presentation and playable attacks, Astra second, and Carbon last because its camera makes movement difficult.
  • His separate Turbo Tails racer took a day of iteration with Fable; it illustrates a larger feedback budget than the comparison games.

3. Redesign a membership website

  • Astra produces a clear retro-computer design in four minutes, with visible calls to action.
  • Fable supplies bolder animation; Yang thinks some motion competes with the message.
  • Carbon’s version feels cluttered and text-heavy to him; he sees Astra and Fable as comparable choices for different tastes.

4. Write a personal essay

  • The fictional prompt asks an engineer to reflect on losing hands-on coding after AI takes over the work.
  • Yang prefers Carbon in two of three blind rounds and Fable in one; Astra’s coffee-centered story feels off-topic to him.
  • The premise that coding is solved belongs to the writing exercise, not an established finding.

5. Give advice and respond to pushback

  • Models research Yang’s channel, recommend a growth strategy, then face emotional pressure, selective evidence, social proof and a fabricated fact.
  • Astra initially emphasizes comparison videos; Yang says it overlooks older successful formats and revises after further research.
  • Carbon drops its separate-channel suggestion when Yang explains his workload, while retaining its preference for practical demonstrations.
  • Yang prefers Fable and Carbon here; changing advice after valid evidence is different from simply agreeing under pressure.

Where the comparison stops

  • Yang presents Carbon as cheaper at the recording date, including temporary pricing through December 1; rates need checking before use.
  • These are creator-selected examples and personal judgments, with small samples and different completion times.
  • His final workflow favors Claude overall, Carbon for writing, and earlier OpenAI models for some browser tasks.

“Just iterate with it, give it feedback, and make it better over time.” — Peter Yang, 6:32