GPT-5.3-Codex Launches: OpenAI’s Fastest AI Coding Agent Sets New Benchmark Records

OpenAI has announced GPT-5.3-Codex, its most advanced code-focused agent to date. According to the company, the new model is 25% faster than GPT-5.2-Codex and has achieved record-level accuracy on the SWE-bench Pro and Terminal-Bench 2.0 benchmarks. Designed for professional development workflows, the system goes beyond suggesting code by handling end-to-end engineering tasks across environments.

On SWE-bench Pro (Public), which evaluates software engineering performance across multiple programming languages, GPT-5.3-Codex reached 56.8% accuracy. The most notable improvement appeared in Terminal-Bench 2.0, focused on terminal command execution, where performance increased from 64.0% in the previous version to 77.3%. The model also showed strong results on OSWorld-Verified, a benchmark that measures how well agents use computer vision to complete desktop tasks. GPT-5.3-Codex scored 64.7%, approaching the human average of 72% and significantly exceeding the previous generation’s 38.2%.

A new “guidance” capability has been introduced in the Codex app, allowing developers to interact with the model while it performs complex operations. This enables real-time adjustments, discussions, and collaborative problem-solving without losing context during code generation or debugging. Training and deployment run on NVIDIA GB200 NVL72 systems, reflecting a co-design effort between OpenAI and NVIDIA to optimize inference performance and reduce token usage during complex tasks.

OpenAI also classified the model as “High Capability” in biosafety and cybersecurity tasks within its Preparedness Framework. GPT-5.3-Codex was specifically trained to identify software vulnerabilities, prompting the implementation of enhanced automated monitoring and controlled access for defensive research.

This release reflects a broader shift from code copilots toward autonomous engineering agents, combining lower latency with improved multilingual workflows. Reports have also suggested future projects involving biometric identity systems to limit bot activity on potential social platforms.

Filed in Robots. Read more about AI (Artificial Intelligence), ChatGPT and OpenAI.