OpenAI’s engineering team shipped the Sora Android app from prototype to global launch in 28 days, working side-by-side with Codex. The four-engineer team consumed ~5 billion tokens, hit a 99.9% crash-free rate, and reached internal release in 18 days—then launched publicly ten days later. Here’s the playbook you can reuse.

Key points

• Radical speed: From zero to production Android app in 28 days with Codex as a “senior engineer” multiplier.

• Quality at pace: 99.9% crash-free at launch; breadth of unit tests generated with Codex.

• Lean team, big surface: 4 engineers orchestrating parallel Codex sessions for features, tests, and refactors.

How it works

OpenAI treated Codex like a capable new teammate: document the architecture, set invariants, and let Codex fill large volumes of code within those guardrails. They established patterns (navigation, DI, networking), wrote a few exemplar features, and used AGENT.md files so Codex consistently followed team standards.

Practical steps (the 28-day plan you can replicate)

Week 1 — Foundations & guardrails

• Pick the Android stack: Kotlin + Jetpack Compose, Hilt for DI, Coroutines/Flows, Paging/WorkManager as needed. Set up Detekt and CI. Create AGENT.md with formatting rules and architecture guidelines for Codex.

• Implement auth + base networking by hand; ship a skeleton feature end-to-end for Codex to mimic.

Week 2 — Parallelise with Codex

• Spin up multiple Codex sessions (playback, search, error handling, tests). Drive each with a short design/implementation plan Codex helps draft; review diffs against the plan.

• Ask Codex to generate broad unit-test coverage and to propose failure points (e.g., player memory optimisations).

Week 3 — Internal release hardening

• Cut an internal testing build via Play Console (up to 100 testers). Integrate Play Integrity API checks for abuse mitigation.

• Use Codex to triage CI failures (paste logs) and produce targeted fixes; enforce crash-free and startup KPIs.

Week 4 — Public launch

• Localise top markets, prepare store listing, and harden privacy screens. Keep Codex running “unsupervised” for long tasks (plans saved to files to overcome context limits).

• Graduate through closed/open tracks quickly if metrics hold; monitor crash-free rate post-launch.

Stack notes: Jetpack Compose is Android’s recommended modern UI toolkit (faster UI delivery), while Play Integrity protects critical flows. Use internal test tracks to ship safely at speed.

FAQs

How did Codex aid development?
Codex read large codebases, drafted implementation plans, generated breadth-first tests, ran in parallel “like multiple new hires,” and applied fixes directly from CI logs—while engineers owned architecture and product quality. OpenAI

Why Android—and what about iOS?
Sora already existed on iOS; Android brought reach. The team often pointed Codex at iOS/back-end repos to mirror flows—“the future of cross-platform is just Codex.” OpenAI

What challenges cropped up?
Left unguided, Codex could drift on architecture (e.g., leak logic into the UI layer). The remedy was robust patterns, exemplar features, and AGENT.md guidance. OpenAI

How ‘real’ is 28 days?
An internal build went live in 18 days, then 10 days to public launch—driven by a tight team and high-leverage use of Codex. OpenAI

Is Sora on Android now?
Yes—OpenAI’s Sora 2 announcement directs users to the Sora app; the Android launch followed the iOS debut. OpenAI

Summary

OpenAI’s approach shows how to pair a lean team with a capable coding agent to ship complex Android apps in weeks—not months. The pattern: set strong foundations, document invariants, plan with the agent, run parallel sessions, and keep humans on architecture and UX. Apply this playbook to your next AI product launch.