There's been a lot of buzz about AI models creating arcade games—especially with the release of Grok 3.
But let's be real: simpler arcade-style games have been within the capabilities of AI models for a while now.
The real challenge is creating a compelling 3D experience that blends engaging gameplay and appealing aesthetics.
To explore this, we made a test comparing Grok 3 against other leading AI models in coding a prototype 3D game, and evaluated it within Rosebud's AI game creation platform.
The AI Game Challenge
Each AI model received the same prompt:
Prompt: "Create a 3D game inspired by Dune, complete with gameplay, crafting, and survival mechanics."
Here's how the AI models stacked up:
Top Performers
- O1 (OpenAI): Struck the best balance between functionality and aesthetics. The generated game looked visually appealing, featured functional controls, and included a basic spice-mining mechanic.
- Grok 3 (xAI): Led the pack in visual and thematic appeal, creating an impressive Dune-inspired spice mining vehicle. However, it fell short on core gameplay mechanics; players could move around but lacked deeper interaction or game objectives.
- Sonnet 3.5 (Anthropic): Delivered a strong overall impression, with solid gameplay mechanics and decent aesthetics, though slightly behind Grok 3 in capturing the atmospheric vibes. It was practical and playable.
Underperformers
- DeepMind Gemini 2.0 Flash & Meta LLaMA: Unfortunately, these models struggled. Despite multiple prompts, they couldn’t produce a playable or visually appealing experience, landing them at the bottom of this particular test.
- DeepMind Gemini 2.0 Flash: Generated an overly difficult game, nearly impossible to play, and visually unattractive.
Challenges and Limitations
A significant challenge with these large language models is their lack of continuity.
Each new prompt generates an entirely fresh game concept, making incremental improvements difficult.
Rosebud’s platform, in contrast, supports continuous iteration through project diffs, enabling creators to refine their games, deploy them instantly, and easily share or remix projects with others.
Conclusion: Best Models for AI Game Prototyping
While Grok 3 captured the aesthetics superbly, o1 delivered the most balanced, functional prototype, making it the standout choice for creators looking for practicality and polish. Sonnet 3.5 also performed well, demonstrating solid potential.
This single test illustrates that, although AI can indeed rapidly prototype compelling 3D games, the choice of model significantly impacts the results.
For game developers exploring AI tools, Rosebud offers a robust environment to make, refine, and share AI-generated projects effortlessly.
Join our Discord channel for more tests, game-making, and community hangouts every Monday and Thursday on Rosebud live streams!