OpenAI has launched Codex in preview. It’s a specialised coding AI agent that allows experienced developers to delegate routine, relatively simple programming tasks to an agent that generates production-ready code and showcases its progress along the way.
Codex is based on the new Codex-1 model, a custom version of OpenAI o3, trained using reinforcement learning for real-world programming tasks.
Unlike other software engineering agents, it performs assigned tasks independently in a separate, isolated environment preloaded with the assigned codebase. This allows it to perform various tasks in parallel, such as writing new features, analyzing the codebase to answer queries, fixing existing bugs, or proposing pull requests for review.
Each assigned task is executed in a specific container preloaded with the user’s codebase and designed to accurately reflect their development environment. To make Codex more effective, developers can include an “AGENTS.md” file in the repository with customised instructions—for example, to contextualise and explain the codebase or to communicate standardisations and style practices for the project.
It’s similar to a README.md, but designed for AI agents instead of humans.
OpenAI claims that Codex typically takes between 1 and 30 minutes to complete tasks, depending on complexity. Once completed, it can also provide evidence, such as logs, for verification. If necessary, users can monitor its progress on each task. OpenAI notes that this tool cannot replace human review:
“It still remains essential for users to manually review and validate all agent-generated code before integration and execution.”
Availability and Pricing
Users can access Codex from ChatGPT and assign software engineering tasks to it by clicking “Code” after entering a request. Currently, it is available only to ChatGPT Pro, Enterprise, and Team subscribers. Plus and Education versions will be supported at a later date.
Since Codex is in its preliminary research phase, OpenAI will offer free access at no additional cost—at least for the next few weeks. Later, OpenAI plans to implement usage-limited access and flexible pricing options for this software engineering AI agent.