OpenAI has announced GPT-5 Codex, a specialized version of its GPT-5 model fine-tuned for coding agent workflows. The new model is designed to support developers by thinking flexibly, working interactively on small tasks, and operating independently on complex jobs.
The model’s training was based on real-world software engineering tasks, including starting projects, refactoring large codebases, improving documentation, debugging, and conducting full code reviews. A key innovation is its ability to dynamically allocate “thinking time” based on a task’s complexity, providing quick responses for simple requests and dedicating more reasoning and iteration for difficult jobs.
Internal tests showed the model could work autonomously for more than seven continuous hours, during which it refined solutions and addressed test failures. The model also demonstrated enhanced capabilities in code reviews by more reliably assessing a pull request’s intent, tracing dependencies, and flagging critical issues.
New features in the Codex development environment
OpenAI recommends using GPT-5 Codex within its specialized environments to maximize functionality. The platform introduces new features and upgrades for the command-line interface (CLI), integrated development environment (IDE), and cloud integration.
CLI and Terminal enhancements
The command-line interface now includes several new features:
- Image attachment: Users can attach screenshots and diagrams to preserve context in discussions.
- Progress tracking: A built-in to-do list helps with task management.
- External tool integration: Improved connections to web search and other platforms.
The terminal UI has also been improved with better formatting for diffs and tool outputs, and simplified access modes—read-only, auto, and full—give developers granular control over the agent’s autonomy.
IDE and Cloud integration
A new IDE extension integrates directly with environments like VS Code, allowing for context-aware interactions. The system can see open files and current selections to generate faster, more precise results.
Codex Cloud integration has been enhanced to significantly reduce task start times. The system can now automatically detect setup scripts, configure dependencies, execute installation commands, and generate screenshots of web UI progress to attach to tasks or pull requests.
Safety and control
Safety remains a primary focus for OpenAI. GPT-5 Codex runs sandboxed by default with no network access and requires explicit permission for potentially risky commands. Developers can customize security policies, including network access limits and approval requirements. Each task generates comprehensive logs, test outputs, and citations to allow for human validation of the agent’s work.
Pricing and availability
Access to GPT-5 Codex is bundled with various ChatGPT subscription tiers, including Plus, Pro, EDU, Business, and Enterprise, with usage limits depending on the selected plan. The model is also available through the Responses API at the same cost as the standard GPT-5.
For developers, GPT-5 Codex acts as an autonomous, context-aware collaborator, using upgraded tools and strong safety constraints to streamline workflows and improve early bug detection.