OpenAI claims it built and launched the Sora Android app in just 28 days, primarily leveraging its AI coding agent, Codex.
The initial production version of Sora was developed between October 8 and November 5, 2025, by a team of four engineers working in tandem with Codex, consuming approximately "5 billion tokens" during development.
The app launched publicly in November and quickly rose to the top of the Google Play Store on its first day. Android users generated "over one million videos" within the first 24 hours.
Engineers Patrick Hum and RJ Marsan explained that the team deliberately avoided scaling up headcount despite tight deadlines, citing Fred Brooks’ well-known principle: “Adding more people to a late software project makes it later.”
Instead, each engineer partnered with Codex to multiply their output. “We formed a strong team of four engineers—each equipped with Codex—to dramatically amplify individual impact,” they stated.
According to OpenAI, Codex ran on an early iteration of the GPT-5.1-Codex model and was responsible for roughly 85% of the codebase. The company has now made this model available to developers via CLI, IDE extensions, and web applications.
Despite the accelerated timeline, OpenAI reports the app maintained a "99.9%" crash-free rate.
The engineering team described their use of Codex as akin to onboarding a "new senior engineer," allowing human developers to focus on architecture, system design, and user experience rather than low-level implementation.
Codex demonstrated strength in reading large codebases, translating logic across platforms, and generating comprehensive test coverage. As noted in the blog: “Codex is (very) enthusiastic about writing unit tests,” and engineers frequently pasted CI logs into prompts to debug failures.
However, the company acknowledged limitations. Codex “still struggles to infer information it hasn’t been explicitly given” and falters at “deep architectural decisions” without clear guidance.
To mitigate this, the team invested heavily in documentation such as AGENTS.md, enforcing coding patterns, standards, and tooling requirements.
A notable technical approach involved using Codex as a cross-platform translation layer instead of relying on shared frameworks. “Forget React Native or Flutter; the future of cross-platform is Codex,” engineers wrote, explaining how Codex translated Swift logic from iOS apps into Kotlin while preserving behavior.
As development sped up, the bottleneck shifted from writing code to reviewing and coordinating parallel Codex sessions. “Our development constraint moved from code generation to decision-making, feedback loops, and integrating changes,” OpenAI said.
In its recently released “State of Enterprise AI 2025” report, the company revealed that Codex’s weekly active users tripled over the past six weeks, accompanied by a ~50% increase in messages processed.
Last October, CEO Sam Altman shared: “Virtually all new code at OpenAI today is written by Codex users.” He added that engineers are completing 70% more pull requests (PRs) per week using Codex.
