Codex

Codex is OpenAI's agentic coding product that executes software engineering tasks asynchronously in cloud-hosted sandboxes. Unlike interactive coding assistants, Codex is designed to handle tasks in the background - a developer assigns a task like implementing a feature, fixing a bug, or writing tests, and Codex works on it independently, returning results with a pull request or code diff for review.

The product evolved from OpenAI's earlier Codex model (which powered the original GitHub Copilot) into a full agentic system built on GPT-5.3-Codex, a model specifically optimized for code generation and software engineering tasks. Each task runs in its own isolated sandbox environment with access to the project's codebase, dependencies, and test suite. Multiple tasks can run in parallel, allowing developers to delegate several pieces of work simultaneously.

Codex represents a different philosophy from interactive tools like Claude Code. Where Claude Code emphasizes real-time collaboration with the developer in the terminal, Codex emphasizes delegation and asynchronous execution. The developer acts more as a project manager, assigning work and reviewing output, rather than pair-programming with the AI. GPT-5.3-Codex leads on specialized coding benchmarks like Terminal-Bench and SWE-Bench Pro but does not publish scores on general reasoning benchmarks, reflecting its focused positioning as a coding-specific agent rather than a general-purpose model.