OpenAI Launches Codex, New AI Coding Agent for Developers

San Francisco, CA — OpenAI has launched a research preview of Codex, its latest AI-powered coding agent, on May 16, 2025. Codex is built on the codex-1 model, optimized specifically for software engineering tasks.

Codex claims to create ‘cleaner’ code than its predecessor, o3, and accurately follows instructions. Moreover, it can iteratively run tests on its output until it yields successful results. The tool is available to subscribers of ChatGPT Pro, Enterprise, and Team starting today.

This AI coding agent operates within a sandboxed, cloud-based environment. Users can load their code repositories and interact with Codex through prompts and buttons in ChatGPT’s sidebar. OpenAI estimates that Codex can complete tasks such as writing simple features, fixing bugs, and running tests within one to 30 minutes.

In response to the rising popularity of AI tools among software engineers, OpenAI plans to introduce rate limits for Codex in the coming weeks. Users will have the option to purchase additional credits. Access to Codex is also expected to expand to ChatGPT Plus and Edu users soon.

OpenAI aims to enhance its suite of AI coding tools further by potentially acquiring Windsurf, a competitor in the AI coding space, for $3 billion. This strategy illustrates OpenAI’s ambition to develop its own capabilities while also making strategic acquisitions.

Josh Tobin, OpenAI’s Agents Research Lead, envisions Codex as a ‘virtual teammate’ that can independently perform tasks typically requiring human engineers, which can take hours or even days to complete. At the same time, OpenAI has emphasized the safety of Codex, ensuring it reliably refuses requests to create malicious software.

The launch of Codex highlights the growing intersection of artificial intelligence and software development, as major tech companies acknowledge that a substantial portion of their source code is now being generated by AI.