OpenAI has introduced GPT-5.3-Codex, a new Codex model positioned as a step beyond “code helper” to a more agentic system that can handle long-running, multi-step work on a computer. The company says GPT-5.3-Codex combines the frontier coding strengths of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2, and that it runs about 25% faster for Codex users.
A central theme of the release is interactivity and steerability on tasks that don’t fit into a single prompt-and-response cycle. OpenAI describes GPT-5.3-Codex as built to handle extended workflows, including debugging, implementing product changes across a codebase, writing and running tests, managing deployments, and iterating through multiple rounds of fixes and improvements while staying aligned with a developer’s guidance.
OpenAI also highlights improvements across both coding and broader “computer use” evaluations. In results, the company reports with “xhigh” reasoning effort, GPT-5.3-Codex scored 56.8% on SWE-Bench Pro (Public), 77.3% on Terminal-Bench 2.0, 64.7% on OSWorld-Verified, and achieved 70.9% wins-or-ties on GDPval. The company additionally claims the model produces more complete, production-ready defaults from simpler or underspecified prompts, with particular gains in web development and iterative project work.
One of the more notable claims in the announcement is that GPT-5.3-Codex was “instrumental in creating itself.” OpenAI says early versions of the model were used internally to debug training, manage deployment, and analyze evaluation results, which the company frames as accelerating the development loop. OpenAI cites examples such as identifying context-rendering issues, investigating caching problems that affect performance, and dynamically scaling GPU capacity to handle traffic surges and stabilize latency during launch periods.
On safety, OpenAI says GPT-5.3-Codex is the first model it classifies as “High capability” for cybersecurity-related tasks under its Preparedness Framework, and the first it has trained directly to identify software vulnerabilities. The company says it is taking a more precautionary approach as capability increases, stating it does not claim definitive evidence that the model can autonomously execute end-to-end cyberattacks, while also describing what it calls its most comprehensive cybersecurity safety stack to date. OpenAI also says it is launching a pilot program, Trusted Access for Cyber, to accelerate defensive security research.
OpenAI says GPT-5.3-Codex is available on paid ChatGPT plans across Codex surfaces, including the app, CLI, IDE extension, and web, and that API access is planned for “soon.” The company also notes the model was co-designed, trained, and served on NVIDIA GB200 NVL72 systems as part of its infrastructure partnership work.
