Jonathan De Vita: OpenAI Launches Codex, a New AI Coding Agent
Jonathan De Vita is a computer scientist who graduated from Lancaster University, having studied AI and coding as part of his degree. This article will look at Codex, a new AI coding agent from OpenAI that is on course to revolutionise the field of software engineering.
In May 2025, OpenAI, the company behind ChatGPT, made a big push in one of the most popular AI domains, unveiling a research preview of Codex, a solution developed for one of the most in-demand areas for AI tools.
Built on a model called codex-1, Codex is capable of performing several tasks simultaneously, including fixing bugs, running tests, writing code and answering questions about a customer’s codebase, according to OpenAI. The codex-1 operating model is a version of OpenAI’s 03 reasoning model created for ChatGPT Pro, Enterprise and Team users.
Pledging to double down on coding, an increasingly in-demand area that has seen rising investment from the likes of Google, Amazon, Microsoft and Anthropic, as well as start-ups like Anysphere, OpenAI conceded that while there were a lot of good AI systems out there, Codex will transform the way developers work by enabling them to delegate more tasks. While different models have different strengths in terms of coding, with Sonnet demonstrating some staying power, things are changing so quickly it is too early to declare a winner, a representative for OpenAI suggested.
Since its launch, OpenAI’s GPT-4.1 model has rapidly gained traction, ranking as the best-performing non-reasoning coding platform in comparison tests. Nevertheless, with many of today’s coding tools operating in tandem in real time with developers, Codex presents revolutionary potential, working on its own in the cloud and delivering an output in one to 30 minutes. In addition to helping with code reviews, OpenAI suggests that the tool can do a better job than its market rivals in inferring an organisation’s coding style.
Unveiling Codex, OpenAI vice president of engineering Srinavas Narayanan highlighted the advanced coding tool as a fundamentally new way of working. He explained that OpenAI would be collecting feedback throughout the research preview, paving the way for a gradual introduction of technologies to advance research, mitigate risks and promote user understanding. Although Mr Narayanan conceded that OpenAI was still in the early phase of product development with much to learn, Codex is positioned as having game-changing potential for coders, with the cloud-based software engineering agent capable of working on multiple tasks in parallel. The software can perform tasks such as writing features, fixing bugs and answering questions regarding the codebase, as well as proposing pull requests for review. Each task runs on the platform’s own cloud sandbox environment preloaded with the user’s repository.
OpenAI’s codex-1 model has been optimised for software engineering, having been trained using reinforcement learning on real-world tasks in a range of different environments. It generates code that closely mirrors PR preferences and human style, adhering precisely to instructions and iteratively running tests until it receives satisfactory results.
Already available to ChatGPT users, Codex is accessed through the sidebar. The user assigns new coding tasks by simply typing a prompt and clicking ‘Code’. Alternatively, if they wish to ask a question about the codebase, the user clicks ‘Ask’. Each task is processed individually in a separate environment preloaded with the codebase.
Capable of reading and editing files and running commands such as text harnesses, linters and type checkers, Codex typically completes tasks in anywhere between one and 30 minutes, depending on complexity, enabling users to monitor its real-time progress. Once the task is complete, the software saves the changes in its environment, providing verifiable evidence of its actions through citations of test outputs and terminal logs to allow the user to trace each step – from the task’s inception to completion.