I asked Claude Code, Codex, and Antigravity to build the same Android app, and only one made something I'd actually install

The test involved challenging each AI agent to autonomously create a detailed Android app, "GearVault," from a single prompt. The app was specified to include a Material 3 interface, gadget inventory, warranty management, local storage, and notifications. The agents were given creative freedom over layout and design to assess their independent development capabilities.

OpenAI Codex produced a functional app with a pleasant interface, scoring 7 out of 10. It handled dark themes well, but its "Vault section" layout was basic, and the app icon lacked consistency with standard Android aesthetics, suggesting a need for more refined detail. Google Antigravity, while initially appearing clean, suffered from critical reliability issues, with non-functional features and bugs emerging during use, rendering it impractical.

Claude Code distinguished itself by creating a visually appealing "Vault section" with a premium interface and notable attention to smaller design details. This outcome suggests that while basic functionality is increasingly achievable across different models, the ability to deliver a polished, user-ready application with thoughtful design remains a key differentiator among current AI coding agents.

This one-shot experiment highlights the varying maturity levels of AI tools in generating not just working code, but also a coherent and reliable user experience. It suggests that developers leveraging these tools may need to factor in significant post-generation refinement depending on the AI agent chosen, especially for consumer-facing applications where design and stability are paramount.