Video summary

ChatGPT vs GLM 5.2 vs Claude Make GTA 6

Main summary

Key takeaways

Gaming

Storyline

  • The video centers on an AI “competition” where multiple models attempt to generate a playable GTA 6–style open-world game from scratch, based on Miami.
  • Each AI outputs a prototype that includes at least a mission segment (e.g., “take down marked hostile”).
  • The main focus is less on narrative depth and more on whether the result is playable and visually impressive.

Gameplay / Build Highlights (what the prototypes allowed)

  • Open-world / mission gameplay: All attempts aim to create a GTA-like environment with missions.
  • Third-person character control: The builds include a character model and basic movement/interaction systems.
  • Weather and time changes (from the ChatGPT build):
    • Weather switching via a key (e.g., T to change weather, with overcast tested).
  • Vehicle interaction:
    • Multiple builds attempted basic car enter/exit functionality, though controls often broke or vehicle entry failed.

Strategies / Key Tips Mentioned (prompting & tooling)

  • The creator used a “basic prompt” to generate GTA 6 “from complete zero,” requesting realism and optimization for the user’s PC.
  • They improved the prompt by iterating with the AI itself:
    • After an initial prompt, they asked the model to rewrite/improve the prompt, then re-pasted it back in.
  • Game engine emphasis (Godot):
    • The AIs were pushed to build using the Godot engine, since GTA-style games are graphics-intensive and Godot was expected to improve rendering/visual quality.
  • Model-specific prompt constraints (for Claude):
    • Claude was instructed to:
      • create a new folder
      • avoid referencing other files
      • not take inspiration from existing assets/files

Results / Comparison Outcomes

ChatGPT (Codex / “5.5 extra high”)

  • Pros:
    • Better-looking graphics
    • Eventually improved control responsiveness after multiple iterations
  • Cons:
    • Initial releases had seriously bugged controls
    • Movement/interaction issues
    • Sometimes couldn’t get into cars even when keys were pressed

GLM 5.2

  • Pros:
    • Eventually got further into a working state (input worked after fixes)
  • Cons:
    • Often struggled with core gameplay functionality
    • Movement/input issues
    • Teleporting problems
    • Car/driving inconsistencies (e.g., animation plays but vehicle entry doesn’t work properly)
  • Overall performance:
    • Seen as weaker/more difficult to stabilize than ChatGPT

Claude (Opus 4.8 / “Ultra code”)

  • Pros:
    • Produced the best overall GTA-like clone in the video
    • Notably strong lighting and reflections
    • Day/night cycle visuals (day → dusk → night transitions)
    • Vehicles “work, kind of,” with generally better interaction than the others
  • Cons:
    • Limited ability to move forward in testing (as mentioned by the creator)
    • Despite that, it still beat the other prototypes in overall quality

End-of-video: Featured gamers/sources

  • Gamers: None mentioned.
  • Sources / systems featured:
    • ChatGPT (Codex 5.5 / “extra high thinking”)
    • GLM 5.2 (Chinese model)
    • Claude (Opus 4.8 / “Ultra code”)
    • Godot engine (used as the target engine in all builds)

Original video