OpenAI Sora for Android: Built in 28 Days Using Codex

OpenAI developed the production version of the Sora Android app in 28 days, moving from prototype to global launch using a lean team of four engineers and the GPT-5.1-Codex model. This rapid development cycle demonstrated that AI-assisted engineering can drastically increase individual impact while maintaining high reliability, resulting in an app with a 99.9% crash-free rate.

AI-Driven Development Workflow

OpenAI treated Codex as a "newly hired senior engineer," focusing on directing and reviewing code rather than manual writing. The team found that while Codex excels at heavy lifting within bounded scopes, it requires human guidance for high-level architectural decisions and user experience.

Capabilities and Strengths of Codex

  • Rapid Codebase Analysis: Codex can quickly read and understand large codebases across multiple programming languages.
  • Testing Coverage: The model is highly effective at writing a broad variety of unit tests to prevent regressions.
  • Feedback Integration: Codex can propose fixes rapidly when provided with CI failure logs.
  • Parallel Execution: The team ran multiple Codex sessions in parallel to handle different modules (e.g., playback, search, error handling) simultaneously.
  • Research and Optimization: Codex was used to sift through SDKs to propose memory optimizations for the video player, minimizing the final app's memory footprint.

Limitations and Human Requirements

  • Lack of Intuition: Codex cannot infer unstated preferences, product strategy, or internal norms.
  • Experiential Gaps: The model cannot run the app to sense if a user flow is confusing or if a scroll feels "off."
  • Architectural Judgment: Left alone, Codex prioritizes immediate functionality over long-term cleanliness, potentially introducing unnecessary view models or misplaced logic.

Technical Implementation Strategy

To ensure the app remained maintainable, the engineering team focused on establishing a rigid foundation before leveraging AI generation.

Manual Foundation and Pattern Setting

Engineers manually implemented the systems design, including architecture, modularization, dependency injection, navigation, authentication, and base networking flows. They then wrote several representative features end-to-end to serve as "correct" examples. This approach ensured that approximately 85% of the code written by Codex followed established team standards and avoided costly refactoring.

The Planning Loop

For non-trivial changes, the team implemented a multi-step planning workflow to enable Codex to work unsupervised for longer periods:

  1. System Understanding: Codex was asked to summarize how a feature works (e.g., data flow from API to UI).
  2. Refinement: Humans corrected the model's understanding of abstractions and layers.
  3. Implementation Planning: Codex created a miniature design document detailing file changes and state introductions.
  4. Execution: Codex applied the plan step-by-step, saving the plan to a file to maintain consistency across context window limits.

Cross-Platform Translation via AI

OpenAI leveraged the existing Sora iOS app as a primary source of truth for the Android version. Rather than using a shared abstraction framework like Flutter or React Native, the team used Codex to translate logic from Swift to Kotlin.

  • Logic Portability: The team utilized the principle that application logic (data models, network calls, validation rules) remains constant across platforms.
  • Contextual Prompting: Codex was prompted to read iOS models and endpoints and propose equivalent Android implementations using the existing API client and model classes.
  • Cross-Repo Navigation: The team used ~/.codex/AGENTS.md files to help Codex discover and navigate between the iOS, backend, and Android repositories.

Implications for Software Engineering

The project shifted the development bottleneck from writing code to making decisions, providing feedback, and integrating changes. OpenAI concludes that AI-assisted development increases the need for human rigor in systems understanding and architecture, as the AI's primary objective is to reach a functional goal quickly. The role of the software engineer is evolving toward a "conductor" who manages AI agents to handle boilerplate and mundane tasks, allowing humans to focus on scalable systems and complex algorithms.

Sources