GPT-5.3-Codex was built to move beyond writing and reviewing code—its design intent centers on true agentic capability across professional computer work. The model merges the frontier coding performance of its predecessor with the reasoning and professional knowledge of the main GPT-5.2 release, creating a system that can take on long-running tasks involving research, tool use, and complex execution. Its defining feature is midtask steering: unlike earlier models that required a prompt-wait-restart loop, Codex lets users steer and interact while it's actively working, without losing context. This shift from a code-writing tool to an autonomous agent that can do nearly anything developers and professionals can do on a computer marks a significant evolution in what AI-assisted development looks like in practice.

The training lineage of GPT-5.3-Codex is uniquely self-referential—it is the first model that was instrumental in creating itself. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations, a process the team described as accelerating its own development significantly. On benchmark performance, the model set a new industry high on SWE-Bench Pro and Terminal-Bench, and achieved 64% on OSWorld-Verified, demonstrating strong real-world agentic capabilities. Available through ChatGPT subscriptions, GitHub Copilot, the Codex macOS app, and across CLI, web, and IDE surfaces, the model brings specialized coding and execution capability directly into the developer workflows where it is needed most.