OpenAI has launched a research preview of Codex, a cloud-based software engineering agent designed to support developers by automating coding tasks such as writing features, fixing bugs, running tests, and suggesting pull requests. Codex operates within secure, isolated cloud environments tailored to the user’s codebase, enabling multiple tasks to run in parallel. It is powered by Codex-1, a specialized version of the o3 model trained to generate human-like code and follow detailed instructions. The tool is now available to ChatGPT Pro, Enterprise, and Team users, with broader access for Plus and Edu users planned soon.
Developers interact with Codex through ChatGPT’s sidebar or Codex CLI, with tasks taking from one to 30 minutes. Each task includes traceable evidence, such as logs and test results, ensuring transparency. Codex reads and modifies code, runs test frameworks, and adheres to project-specific guidelines using AGENTS.md files. Safety remains a priority, with strict boundaries to prevent misuse. OpenAI has worked with early testers like Cisco and Superhuman to fine-tune performance across diverse environments. Codex helps reduce repetitive work, allowing developers to focus on core challenges while maintaining control over the final output. Additional models, including codex-mini-latest, are available for local workflows through the Codex CLI.
