Codex / OpenAI, YouTube
OpenAI has announced the launch of a research preview version of Codex, an artificial intelligence agent for programming. The tool is based on the OpenAI o3 model.
Codex allows developers to delegate routine and relatively simple programming tasks to an AI agent that will generate ready-to-use code and demonstrate its performance. Its interface can be accessed from the sidebar of the ChatGPT web application. Users enter a query and then click either «code» to start generating code or «ask» to get answers to questions and tips.
Each time Codex follows a task, it executes in a separate pre-loaded container with the user’s codebase that accurately replicates their development environment. To make Codex more efficient, developers can include an AGENTS.md file in the repository with special instructions — something like README.md, but for AI agents.
Codex is built on codex-1, an improved variation of models of reasoning o3 by OpenAI, trained on a wide range of coding tasks. The tool supports code analysis and generation, as well as test iteration in the process.
OpenAI’s official announcement of Codex contains many rebuttals to common arguments against AI-based coding agents. In particular, based on their experience with earlier models, many developers say that LLM coding tools, especially for web coding, produce scripts that are not compliant with standards, opaque or difficult to debug, or unsafe.
Fine-tuning codex-1 should partially solve these problems. Codex is able to show the progress of its own «reasoning» and preliminary results at each stage of the process (which can last from one to 30 minutes). OpenAI notes that it’s still important for users to manually review and validate all agent-generated code before integration and execution.
Codex is available in a trial mode, but it is currently deployed for all users ChatGPT Pro, Enterprise, and Team. Support for Plus and Edu plans will be available later. Users will have access at no additional cost for «the next» weeks so they can explore the tool’s capabilities. However, OpenAI says it will introduce a speed limit and a new pricing scheme at a later date.
Sources: OpenAI, Ars Technica