OpenAI now uses its own AI coding tool to build and improve that same tool. The company’s product lead says most of Codex is built by Codex itself. Engineers at the company rely on the agent daily to write features, fix bugs, and propose code changes.
How Codex Works Inside OpenAI
According to Ars Technica, Alexander Embiricos leads product development for Codex at OpenAI. He said the vast majority of Codex is built by Codex. The tool runs in cloud environments and handles tasks in parallel.
OpenAI launched Codex as a research preview in May 2025. The company offers the tool through ChatGPT’s web interface and command-line tools. Extensions work with VS Code, Cursor, and Windsurf.
The Codex name dates back to a 2021 model based on GPT-3. That earlier version powered GitHub Copilot’s code completion feature. Embiricos said staff believe the name is short for code execution.
Real-World Results at OpenAI
The tool helped build OpenAI’s Sora Android app. Four engineers created the app from scratch in 18 days. The team shipped it to the app store in 28 days total.
Usage jumped after OpenAI released the command-line version. External developer adoption grew 20 times after the CLI launched alongside GPT-5 in August 2025. OpenAI released GPT-5 Codex, a specialized version for coding, on September 15.
Engineers as Teammates
OpenAI treats Codex like a junior team member. The company uses the same open-source version that external developers can download. Employees assign tasks to Codex through project management tools like Linear.
Ed Bayes designs interfaces for Codex. He said the tool integrates with Slack and other platforms. Team members can tag Codex in messages and ask it to fix issues. The agent creates pull requests that humans review.
Codex monitors its own training runs and processes user feedback. Engineers can ask it to look at feedback and decide what to build next. The tool writes research code for its own training.
Independent research shows mixed results. A METR study in July found experienced developers were 19 percent slower when using AI tools on complex codebases. The researchers noted AI may perform better on simpler projects.
Embiricos said OpenAI has not reduced headcount because of Codex. A human stays in the loop to review code. The company sees the tool as an amplifier of human work, not a replacement.
