OpenAI has officially launched Codex, a powerful new AI coding agent designed to transform the way developers write, test, and maintain software. Currently available as a research preview for select ChatGPT subscribers, Codex builds on OpenAI’s vision of seamless human-AI collaboration in the realm of software engineering.
Unlike its predecessor — the original Codex model behind GitHub Copilot — this latest iteration is powered by a fine-tuned version of the OpenAI o3 reasoning model, dubbed Codex-1. Trained via reinforcement learning on a wide range of coding tasks, Codex-1 emphasises cleaner code generation, test-driven development, and enhanced transparency.
Codex doesn’t just churn out lines of code. It tests its own outputs, logs results, and even provides citations from terminal logs and test outputs. It operates in cloud-based sandboxes that mirror a developer’s environment and can connect to GitHub to preload repositories. Once onboarded, Codex can:
Build new features
Fix bugs
Answer technical questions
Run tests
Summarise code changes
Each task runs in a secure, isolated container with detailed logs for developer review. It also supports simultaneous task execution without interrupting the user’s workflow.
To customize Codex’s performance, developers can create an “AGENTS.md” file, similar to a README but tailored for AI — guiding the agent on code style, project norms, and specific requirements.
In a bid to address long-standing concerns about AI misuse, Codex operates in an air-gapped environment, isolated from the broader internet and APIs to prevent unauthorised access or malicious behaviour. OpenAI has also integrated real-time monitoring systems to flag and block harmful requests — though human oversight is still required, as Codex is not immune to occasional hallucinations or errors.
Some of tech’s biggest names are already exploring what Codex can do:
Cisco is testing it to streamline engineering workflows.
Superhuman uses Codex to improve test coverage and enable non-engineers to contribute.
Kodiak relies on it to debug complex codebases used in autonomous vehicles.
Temporal utilises the agent for background tasks like test creation and debugging.
OpenAI’s rollout plan includes ChatGPT Pro, Enterprise, and Team users, with Plus and Edu subscribers expected to gain access soon. For now, the tool is free for those with access, though rate limits and a paid credit system are on the horizon as usage scales.
Despite the hype, Codex still has limitations. It doesn’t support image inputs for frontend development and doesn’t allow user intervention mid-task. And while remote agent execution is more powerful, it's not yet as quick as local development for small edits.
Still, OpenAI is betting big on asynchronous, autonomous agents becoming core to the future of coding. With demand for AI coding assistants booming — and rivals like Google and Anthropic entering the arena — Codex is OpenAI’s high-stakes play to lead the next era of software development.