OpenAI has announced the launch of GPT-5-Codex, an advanced version of its GPT-5 model designed specifically for agentic coding. The company stated that the new system has been developed to improve reliability in handling multi-step programming tasks, refactoring codebases, and conducting automated reviews.
According to OpenAI, GPT-5-Codex builds on previous Codex models with improved steerability, enabling developers to provide high-level instructions that the system can follow across projects. It has been optimised to enforce style and cleanliness guidelines consistently, reducing the need for repeated corrections.
Enhanced Coding and Review Features
GPT-5-Codex has been designed to process both small and large coding tasks. For short prompts, the model responds with reduced latency and fewer tokens, while for complex workflows, it allocates more reasoning time and iterations.
The company highlighted its upgraded review capabilities. GPT-5-Codex can analyse an entire codebase, including dependencies and test suites, to identify bugs, regressions, and test failures. It can also run automated tests to validate outputs. Integration with tools OpenAI has integrated GPT-5-Codex into multiple environments.
In the command-line interface, the model supports progress tracking and an updated terminal interface. In IDEs such as Visual Studio Code and Cursor, it can maintain context across open files, manage local and cloud workspaces, and preview changes before execution.
The model is also able to process screenshots or user interface designs, enabling it to generate or debug code for front-end applications. According to the company, this functionality allows GPT-5-Codex to work effectively on mobile and web projects.
Cloud Setup and Permissions
Cloud workflows have been updated with cached containers that reduce environment setup times. The model can automatically install project dependencies, streamlining configuration.
GPT-5-Codex includes permission modes, ranging from read-only access to full execution rights. By default, it runs in a sandboxed environment without external network access. OpenAI stated that additional safeguards are applied in sensitive use cases, including areas involving biological or chemical applications.
