Today, OpenAI announced GPT-5.3-Codex, its most capable agentic coding model to date. The company says the model achieved record scores on SWE-bench Pro and Terminal-Bench. Along with improved coding performance, OpenAI claims GPT-5.3-Codex is 25% faster than GPT-5.2-Codex.
OpenAI’s Codex-series models are primarily targeted at developers. On the SWE-bench Pro (Public) benchmark, GPT-5.3-Codex scored 56.8% (vs. 56.4% for GPT-5.2-Codex and 55.6% for GPT-5.2). On Terminal-Bench 2.0, it scored 77.3% (vs. 64.0% for GPT-5.2-Codex and 62.2% for GPT-5.2). On the OSWorld-Verified agentic computer-use benchmark, the model scored 64.7% (vs. 38.2% for GPT-5.2-Codex and 37.9% for GPT-5.2).
OpenAI also says GPT-5.3-Codex achieved these results while using fewer tokens than prior Codex models. Thanks to inference-stack improvements, the company claims the model runs 25% faster for Codex users.
OpenAI is also positioning GPT-5.3-Codex as a better collaborator for developers. The company says users can “steer and interact” while the model is working on a task, “without losing context.” In the Codex app, GPT-5.3-Codex provides frequent progress updates as it works, allowing developers to ask questions, discuss approaches, and guide it toward the desired solution in real time.
OpenAI added that early versions of GPT-5.3-Codex performed well enough that the company used them to improve training and support the deployment of later model versions.
OpenAI wrote the following regarding Codex in the announcement blog post:
What started as a focus on being the best coding agent has become the foundation for a more general collaborator on the computer, expanding both who can build and what’s possible with Codex.
GPT-5.3-Codex is now available to all ChatGPT paid plan users. Users can access it via the Codex app, CLI, IDE extension, and the web. OpenAI also said the model will be available via the API soon.
Finally, OpenAI noted that GPT-5.3-Codex was co-designed for, trained with, and served on NVIDIA GB200 NVL72 systems.