OpenAI's GPT-5.3-Codex: The AI That Helped Build Itself

A New Kind of AI Model

OpenAI has launched GPT-5.3-Codex, an advanced coding model that breaks new ground by being the first AI system instrumental in its own development. The company's research team used early versions of the model to debug training runs, manage deployment, and diagnose test results, marking a significant shift in how frontier AI systems are built.

Performance and Benchmarks

GPT-5.3-Codex operates twenty-five percent faster than its predecessor while achieving state-of-the-art results across industry benchmarks. It scored fifty-six point eight percent accuracy on SWE-Bench Pro and seventy-seven point three percent on Terminal-Bench 2.0, representing a thirteen-point jump over the previous model. On OSWorld-Verified, it reached sixty-four point seven percent, approaching human-level performance of around seventy-two percent.

Cybersecurity Milestone

The model is the first from OpenAI to be classified as high capability for cybersecurity tasks under the company's Preparedness Framework. It achieved an eighty percent success rate on internal cyber range evaluations, solving scenarios involving binary exploitation and command-and-control operations. In response, OpenAI is launching Trusted Access for Cyber, a pilot programme for defence research, alongside ten million dollars in API credits for its Cybersecurity Grant Programme.

The Codex Desktop Experience

Alongside the model release, OpenAI introduced a new Codex desktop application for macOS, functioning as a command centre for managing multiple AI coding agents in parallel. The app supports worktrees for conflict-free collaboration and includes Skills, bundled instructions that connect to tools like Figma, Linear, and cloud deployment services. Agents can run independently for up to thirty minutes before returning completed work.

Availability

GPT-5.3-Codex is available now across ChatGPT's paid plans, the Codex desktop app, command-line interface, and IDE extensions, with API access expected to follow.