Need code? Have Codex

Blitz Bureau

OpenAI last week introduced a new AI tool called Codex that is designed to handle multiple software engineering-related tasks at the same time, from generating code for new features to answering questions about a user’s codebase, fixing bugs, and suggesting pull requests for code review

The cloud-based, AI agent-driven coding tool runs these tasks in its own cloud sandbox environment that has been preloaded with a user’s code repository, says a report in The Indian Express.

Codex has been released under research preview. However, all ChatGPT Pro, Enterprise, and Team users have access to the AI coding tool. “Users will have generous access at no additional cost for the coming weeks so you can explore what Codex can do, after which we’ll roll out rate-limited access and flexible pricing options that let you purchase additional usage on-demand,” OpenAI said in a blog post.

ChatGPT Plus and Edu customers will be given access at a later date, the Microsoft-backed AI startup added.
OpenAI’s latest offering comes at a time when AI is poised to disrupt the software engineering sector, raising widespread fears of job displacement. Microsoft CEO Satya Nadella recently said that 30 per cent of the company’s code is now AI-generated. A few weeks later, the tech giant announced it is laying off 6,000 employees or 3 per cent of its workforce, with programmers reportedly being impacted the most.
“It still remains essential for users to manually review and validate all agent-generated code before integration and execution,” OpenAI noted in its Codex announcement blog post.

What is Codex?
With Codex, developers can delegate simple programming tasks to an AI agent. It has its own unique interface that can be accessed from the side bar in the ChatGPT web app.

Codex is powered by codex-1, an AI model that is a variation of OpenAI’s o3 reasoning model. Except that codex-1 has been specifically trained on a wide range of real-world coding tasks to analyse and generate code “that closely mirrors human style and PR preferences, adheres precisely to instructions.”

Its outputs have further been fine-tuned using reinforcement learning so that Codex-1 can “iteratively run tests until it receives a passing result.” In terms of performance and accuracy, OpenAI said that Codex-1 fared better than its o3 AI model when evaluated on its internal SWE benchmark as well as the company’s human-validated version of it (SWE-bench Verified).

How does Codex work?
Codex can read and edit files as well as run commands including test harnesses, linters, and type checkers. It typically takes anywhere between one minute to 30 minutes to complete a task depending on the difficulty level, as per OpenAI.

The AI coding agent performs each task in a distinct, isolated environment that is preloaded with the user’s codebase serving as context. “Like human developers, Codex agents perform best when provided with configured dev environments, reliable testing setups, and clear documentation,” OpenAI said.
Another unique feature of Codex is that it shows its thinking and work with every step as it goes about completing the task(s). In the past, several developers have pointed out that AI coding agents produce coding scripts that do not follow standards and are difficult to debug.