Today, OpenAI released a groundbreaking new cloud-based AI programming agent called Codex during an online live stream, marking the official start of a new era in AI programming. The six-person team led by Greg Brockman showcased this powerful tool to audiences worldwide, demonstrating its ability to complete software engineering tasks that might have previously taken days in just half an hour, significantly boosting development efficiency.
Powered by the new codex-1 model, which is a specialized version of OpenAI's o3 model tailored for software engineering, Codex can safely and efficiently handle multiple tasks in parallel within a cloud sandbox environment. It seamlessly integrates with GitHub, allowing direct access to code repositories. Codex can rapidly build functional modules, deeply answer questions about codebases, accurately fix code bugs, submit pull requests, and automatically run tests for verification. These functions, which might have taken developers hours or even days in the past, can now be accomplished by Codex in up to 30 minutes.
Running on OpenAI’s computational infrastructure, Codex shares the same system as reinforcement learning. Each task runs in an independent virtual sandbox equipped with its own file system, CPU, memory, and network policies, ensuring both efficiency and safety. Codex performs exceptionally well in the preparedness repository and handles the CodeX CLI library effortlessly, showcasing its generalization capabilities across different projects.
To ensure Codex outputs align closely with human coding preferences and standards, OpenAI paid special attention to real-world coding tasks and diverse environments while training codex-1. Benchmark tests show that codex-1 achieved a high score of 72.1% on SWE-bench, outperforming Claude3.7 and o3-high. OpenAI also provided practical comparison examples of Codex and o3 in four open-source libraries (astropy, matplotlib, django, and expensify), showing that the code generated by Codex is more concise and efficient.
Currently, Codex has been officially launched for global ChatGPT Pro, Enterprise, and Team users. Plus and Edu users will soon have access as well.
Meanwhile, OpenAI released a lightweight version of the model optimized for Codex CLI—the o4-mini version of codex-1—and simplified the login process. Developers can now log in directly using their ChatGPT accounts. To encourage usage, users logging into Codex CLI with a ChatGPT account within the next 30 days will receive free credits.
Although Codex is currently in research preview mode and does not yet support front-end capabilities such as image input or the ability to make real-time corrections during task execution, OpenAI plans to enhance its interactivity and flexibility in the future. This includes providing guidance midway through tasks, collaborating with AI to implement strategies, receiving proactive progress updates, and deep integration with commonly used tools. The arrival of Codex could potentially reshape the underlying logic of software development and ignite a programming revolution.