OpenAI has launched Codex, a cloud-based research preview of a software engineering agent designed to automate common development tasks such as writing code, debugging, testing, and generating pull requests. Codex is integrated into the Professional, Team, and Enterprise versions of ChatGPT, with each task running in a secure sandbox environment preloaded with the user's codebase and configured to reflect their development setup.
Codex is powered by codex-1, a version of OpenAI's o3 model optimized for programming tasks. Trained on real-world examples using reinforcement learning, it can generate human-like code. The model repeatedly runs code and tests until it finds the correct solution. Once a task is completed, Codex submits its changes within the sandbox, providing test outputs and terminal logs to ensure transparency.
The Codex sidebar in ChatGPT allows users to assign tasks through text prompts or ask questions about their codebase. The model can edit files, run commands, and execute tests, with typical completion times ranging from one minute to thirty minutes. Codex supports AGENTS.md files—repository-level instructions that help guide agents through project-specific practices and testing procedures.
The Codex CLI is an open-source command-line companion interface that operates using API credits. As Fouad Matin, a technician at OpenAI, clarified, access to Codex within ChatGPT is included in Professional, Team, and Enterprise subscriptions:
Codex is included in the pricing of ChatGPT (Professional, Team, Enterprise) with generous access available for the next two weeks.
However, the system currently does not support full application testing for apps with live user interfaces. As one user pointed out:
Most software engineering today is web development. How does it handle situations where there are separate layers, environment variables, and UI interfaces? Does it actually run the application so users can test it, or do you need to push changes and pull down copies for local testing? That would be very frustrating. Ideally, future agents will be able to test themselves, but I guess they're not quite there yet.
Codex operates in an isolated container without internet access or UI execution capabilities. While it can handle test suites, code linters, and type checkers, final validation and integration remain the responsibility of human developers.
OpenAI has also introduced Codex mini, a lightweight model designed for faster interactions and lower latency, now the default engine in Codex CLI and available via API as codex-mini-latest. It is priced at $1.50 per million input tokens and $6 per million output tokens, with a 75% prompt caching discount.
This release reflects OpenAI’s broader strategy of eventually supporting real-time AI coding assistants and asynchronous agent workflows. While Codex currently connects with GitHub and is accessible from ChatGPT, OpenAI envisions deeper integrations in the future, including support for task assignments from Codex CLI, ChatGPT desktop, and tools like issue trackers or CI systems.
