Today, OpenAI unveiled Codex, a cloud-based AI agent designed for software engineering and powered by their most advanced coding model, codex-1. Starting today, Codex is accessible to ChatGPT Pro, Team, and Enterprise subscribers, with ChatGPT Plus and Education users slated to gain access soon.
Codex can tackle multiple development tasks simultaneously in the cloud: it can implement new features, debug code, answer queries about your codebase, and even generate pull requests for review. By linking their GitHub repositories, developers can invoke the agent directly from ChatGPT’s left sidebar to automate a wide range of workflows.
Built on the o3 architecture, codex-1 was fine-tuned via reinforcement learning on real-world coding challenges across diverse environments, enabling it to write code that aligns with typical human style and pull-request conventions. It carefully follows instructions, running tests in its cloud sandbox until they pass. You can customize its behavior by including an AGENTS.md file in your repo to guide its navigation and actions.
In internal benchmarks, codex-1 achieved 75% accuracy on OpenAI’s proprietary software-engineering tasks—outperforming o3-high’s 70%—and showed marginal gains over o3-high on the SWE-Bench Verified suite after multiple attempts. For security, the Codex agent runs within an isolated container with no internet access, ensuring it cannot reach external sites, APIs, or services during execution.