I Made Codex and Claude Code Build the Same App. One Clearly Won.

Summary

The video presents an experiment comparing two AI coding agents, Claude Code and Codex, by giving them the exact same prompt to build a production ready Typeform alternative. The prompt instructs the agents to orchestrate specialized subagents through research, build, and verify phases, ensuring complete functionality without stopping at prototypes and requiring rigorous testing across user workflows. The creator applies the prompt inside both coding environments to see how the resulting applications and development processes differ.

Claude Code produces an application named Formora, while Codex produces an application named Rillform. Formora features a clean interface that feels deliberately designed and easy for users to navigate. Rillform presents a more cluttered interface with numerous form settings and options visible on the main workspace, which can feel overwhelming to a user. Both applications are tested by filling out forms and reviewing the resulting data dashboards and submission logs, confirming that both successfully collect and display user responses.

Comparing the resource usage and performance metrics reveals stark differences between the two agents. Claude Code completes the build in 5 hours and 32 minutes at a cost of $832, consuming approximately 2.1 million output tokens. It operates with 1 orchestrator, 35 subagents, and 2,824 tool calls. Codex takes 61 hours and 47 minutes at a cost of $2,932, consuming over 11.4 million output tokens. Codex operates with 1 orchestrator, 126 subagents, and 32,561 tool calls. Both systems use GPT-5.6 Soul as their underlying model, making the divergence in time, cost, and token consumption particularly notable.

Testing metrics also highlight different approaches to verification. Claude Code runs 296 unit tests, 199 test cases, and 102 browser tests. Codex runs a much higher volume, including 2,339 unit tests, 341 test cases, and 391 browser tests across multiple device profiles. While Codex executes more extensive testing and builds a robust backend infrastructure with immutable revisions and offline recovery, Claude Code proves significantly faster and more economical while maintaining a superior user experience.

Finally, the creator has Codex evaluate both applications across four categories. Codex scores Claude Code higher in product judgment, scope, pragmatism, and efficiency, while scoring Codex higher in architecture and execution. Ultimately, Codex concludes that Claude Code wins overall, demonstrating that Claude Code delivered a faster, cheaper, and more user focused result despite using fewer resources and running fewer tests.