OpenAI builds Sora Android app in 28 days with Codex assistance
In just under a month, OpenAI turned a fresh concept into a functional Android app for Sora, a move that would normally demand weeks of coordinated engineering. The sprint unfolded across 28 days, a timeline that left little room for the usual back‑and‑forth of code reviews, UI tweaks, and platform testing. To meet that cadence, the team turned to Codex, the company’s code‑generation model, as a core piece of the workflow.
According to the developers, Codex was tasked with parsing extensive codebases and converting existing logic into the new mobile context, while human engineers concentrated on higher‑level decisions—system architecture, user experience, and overall quality. The claim that the AI “excelled at reading large codebases, translating logic” suggests a division of labor where the model handled repetitive, well‑defined tasks, freeing the team to address design and integration challenges. This balance of machine assistance and human oversight frames the authors’ later reflection on how Codex contributed to the project’s speed and scope.
"We leaned on Codex to do a huge amount of heavy lifting inside well-understood patterns and well-bounded scopes, while our team focused on architecture, user experience, systemic changes, and final quality," the authors wrote. OpenAI said Codex excelled at reading large codebases, translating logic across platforms, and generating broad test coverage. "Codex is (uniquely) enthusiastic about writing unit tests," the blog noted, adding that engineers frequently pasted CI logs into prompts to diagnose failures.
Codex "isn't yet great at inferring what it hasn't been told," and struggled with "deep architectural judgment" when left unguided. To address this, the team invested heavily in documentation, such as AGENTS.md, to enforce patterns, coding standards, and tooling requirements. One notable technique was to use Codex as a cross-platform translation layer rather than as a shared framework.
"Forget React Native or Flutter; the future of cross-platform is just Codex," the engineers wrote, explaining that Codex translated Swift logic from the iOS app into Kotlin while preserving behaviour. As development accelerated, the bottleneck shifted from writing code to reviewing and coordinating parallel Codex sessions. "Our bottleneck in development shifted from writing code to making decisions, giving feedback, and integrating changes," OpenAI said.
In the company's State of Enterprise AI 2025 report, released a few days ago, the company stated that over the last six weeks, there was a 2x increase in weekly active Codex users. Further, the company observed a ~50% increase in Codex messages over the same period. In October, Sam Altman, the CEO of the company, revealed that "Almost all new code written at OpenAI today is from Codex users." He added that engineers in OpenAI complete 70% more pull requests (PRs) each week using Codex.
OpenAI’s claim that a four‑engineer team delivered Sora’s Android client in just 28 days, consuming roughly five billion tokens, invites both curiosity and caution. The team attributes most of the heavy lifting to Codex, noting that the AI excelled at parsing large codebases and translating logic within well‑bounded scopes, while developers focused on architecture and user experience. The app’s public launch in November saw it top the Google Play Store on day one, with Android users generating more than a million videos.
Yet the report offers no detail on post‑release stability, long‑term maintenance costs, or how the token consumption translates into development effort. It's unclear whether similar timelines could be achieved on less familiar projects or with smaller token budgets. Consequently, while the headline numbers are striking, the broader implications for software engineering practices are still uncertain.
Some developers question whether the approach scales beyond a tightly scoped prototype. Without independent benchmarks, the claim sits on internal metrics alone.
Further Reading
- Sora 2 is here - OpenAI
- Introducing Codex - OpenAI
- Papers with Code - Latest NLP Research - Papers with Code
- Hugging Face Daily Papers - Hugging Face
- ArXiv CS.CL (Computation and Language) - ArXiv
More in Business & Startups
Runway launches General World Model with three versions and Gen‑4.5 upgrades
Runway just dropped a new AI suite that promises more than a single‑purpose generator.
Disney sues Google over AI “virtual vending machine” after USD 1 bn OpenAI deal
Disney has filed a lawsuit against Google, accusing the tech giant of using its...
NVIDIA invests USD 5 billion in Intel, gaining a significant equity stake
Why does a chipmaker best known for graphics cards suddenly pour billions into a rival’s balance...
Broadcom discloses USD 21 B Anthropic order for Google’s 7th‑gen TPUs
Broadcom’s latest filing reveals a $21 billion purchase from Anthropic, the AI startup that has...
Harness secures USD 240 M E round, led by Goldman Sachs, for AI software delivery
Harness just closed a $240 million Series E, with Goldman Sachs heading the round, to fund its...
Common Questions Answered
How did OpenAI manage to build the Sora Android app in just 28 days?
OpenAI leveraged Codex, its code‑generation model, to handle heavy lifting such as parsing large codebases and translating logic across platforms. This allowed a four‑engineer team to focus on architecture, user experience, and final quality while meeting the tight timeline.
What role did Codex play in generating test coverage for the Sora Android client?
Codex was described as "uniquely enthusiastic about writing unit tests," automatically producing broad test suites that covered many code paths. Engineers also used Codex to interpret CI logs, ensuring that the generated tests aligned with continuous integration requirements.
How many tokens did the team consume while developing the Sora Android app, and why is this significant?
The development process consumed roughly five billion tokens, highlighting the extensive use of Codex for code generation and analysis. This token count underscores both the power of large language models in software engineering and the need for careful resource management.
What was the public reception of the Sora Android app upon its launch in November?
When the app launched publicly in November, it topped the Google Play Store rankings on its first day, indicating strong user interest and rapid adoption. The immediate success suggests that the rapid development approach did not compromise the app's quality or appeal.