OpenAI says GPT-5 Codex now writes most of the code used to improve itself

OpenAI is using its GPT-5-based coding agent, Codex, to generate the bulk of the code the company uses to build and improve the agent itself, employees told Ars Technica. "I think the vast majority of Codex is built by Codex," Alexander Embiricos, product lead for Codex at OpenAI, said.

OpenAI launched the modern Codex as a research preview in May 2025; it runs as a cloud-based software engineering agent that can write features, fix bugs, propose pull requests, and execute tasks in sandboxed environments linked to code repositories. The tool is available through ChatGPT’s web interface, a command-line interface (CLI), and IDE extensions for VS Code, Cursor, and Windsurf.

Embiricos said the team even uses the same open-source CLI that external developers can download and modify. OpenAI staff described Codex monitoring its own training runs, processing user feedback to "decide" what to build next, and being assignable via project-management tools like Linear or communication platforms like Slack.

Embiricos credited Codex with speeding internal projects — citing the Sora Android app, which he said four engineers built in 18 days and shipped to an app store in 28 days — while designer Ed Bayes said Codex functions as a "teammate" that lets non-engineers prototype and contribute code.

Key Topics

Tech, Openai, Codex, Alexander Embiricos, Sora Android, Chatgpt