In November, OpenAI launched the Sora app for Android, allowing users to create videos from prompts. The app quickly hit #1 on the Play Store, with over a million videos made in the first day. The team built the app in just 28 days using Codex, an AI tool designed to assist developers. During the development, the team processed around 5 billion tokens while maintaining a 99.9% crash-free rate. They used an early version of the GPT-5.1-Codex model, which is available for other developers as well. The team approached Codex as if it were a new senior engineer, benefiting from its strengths while providing necessary guidance. Codex struggles with context and deep architectural decisions, often requiring clear instructions on team practices and project goals. To streamline this process, they created AGENT.md files that outlined coding styles and best practices. Codex excelled in quickly understanding code, writing unit tests, and responding to feedback, enabling faster iterations. Despite relying heavily on Codex—estimated at 85% of the codebase—the team emphasized the importance of a solid foundation. They designed the app’s core architecture and implemented key features manually to ensure maintainability. This groundwork allowed Codex to work more independently while adhering to the team’s standards. The goal was to establish clear expectations for how the code should function, ultimately leading to a more cohesive and efficient development process.