According to Inc, OpenAI has been shipping products at a relentless pace in the second half of 2025, releasing multiple AI models, ChatGPT features, an AI-powered browser, and Sora 2. The secret weapon is Codex, OpenAI’s family of agentic coding tools that allow AI agents to write, edit, and run code at impossible human speeds. Codex product lead Alex Embiricos says daily usage jumped 10x since August’s GPT-5 release, with 92% of OpenAI engineers heavily using it internally. The system wrote roughly 85% of the Android version of Sora app and enabled a team of four engineers to function like sixteen by running multiple Codex instances simultaneously. Engineer John Nastos uses Codex for 99% of his ChatGPT Atlas browser code, estimating features take one-fourth the time compared to manual coding.
The Codex reality check
Here’s the thing about AI coding assistants – they’re incredibly powerful, but they’re not magic. The Sora Android app story reveals the current limitations. When engineers first gave Codex the iOS codebase and told it to create the Android version, it worked for 12 hours straight and produced something “certainly wasn’t anything that we could show anybody.” Basically, Codex is like “a senior software engineer that just got hired” – technically skilled but clueless about company-specific best practices and product vision.
The human-AI partnership that actually works
So what’s the winning formula? The Sora team spent their first week writing code by hand to establish architecture and best practices, creating what they call a “context-rich environment.” Then Codex could operate with proper guidance. This pattern repeats across OpenAI – humans set the vision and constraints, then unleash multiple Codex instances to execute in parallel. The result? That four-person team delivered a working Android app in 18 days and launched publicly 10 days later. That’s insane velocity for what would traditionally take months.
What this means for everyone else
Look, if OpenAI is using this to ship faster, you can bet every tech company is racing to adopt similar tools. Anthropic’s Claude Code, Google’s Gemini, and startups like Replit are all in this space. But here’s the interesting part – Embiricos anticipates Codex will become more intuitive for non-engineers. He suggests non-coders start using it for simple side projects instead of bothering engineering teams. Think about that for a second – we might be heading toward a world where technical and non-technical people collaborate through AI intermediaries.
The velocity vs quality question
Now, I’ve got to ask – is faster always better? When Codex can build features in the background while engineers debate whether to include them, does that lead to better products or just more features? The ChatGPT Atlas browser story shows how AI can handle complexity across multiple codebases (Swift, JavaScript, Python simultaneously) that would overwhelm human engineers. But who’s checking the AI’s work? The balance between human oversight and AI autonomy is still being figured out, even at the companies building these tools.
