GPT-5.3-Codex Launches: OpenAI’s Quickest AI Coding Agent Units New Benchmark Information

OpenAI has announced GPT-5.3-Codex, its most superior code-focused agent so far. In accordance with the corporate, the brand new mannequin is 25% quicker than GPT-5.2-Codex and has achieved record-level accuracy on the SWE-bench Professional and Terminal-Bench 2.0 benchmarks. Designed for skilled improvement workflows, the system goes past suggesting code by dealing with end-to-end engineering duties throughout environments.

On SWE-bench Professional (Public), which evaluates software program engineering efficiency throughout a number of programming languages, GPT-5.3-Codex reached 56.8% accuracy. Probably the most notable enchancment appeared in Terminal-Bench 2.0, targeted on terminal command execution, the place efficiency elevated from 64.0% within the earlier model to 77.3%. The mannequin additionally confirmed robust outcomes on OSWorld-Verified, a benchmark that measures how properly brokers use laptop imaginative and prescient to finish desktop duties. GPT-5.3-Codex scored 64.7%, approaching the human common of 72% and considerably exceeding the earlier technology’s 38.2%.

A brand new “steering” functionality has been launched within the Codex app, permitting builders to work together with the mannequin whereas it performs complicated operations. This allows real-time changes, discussions, and collaborative problem-solving with out dropping context throughout code technology or debugging. Coaching and deployment run on NVIDIA GB200 NVL72 methods, reflecting a co-design effort between OpenAI and NVIDIA to optimize inference efficiency and scale back token utilization throughout complicated duties.

OpenAI additionally labeled the mannequin as “Excessive Functionality” in biosafety and cybersecurity duties inside its Preparedness Framework. GPT-5.3-Codex was particularly skilled to determine software program vulnerabilities, prompting the implementation of enhanced automated monitoring and managed entry for defensive analysis.

This launch displays a broader shift from code copilots towards autonomous engineering brokers, combining decrease latency with improved multilingual workflows. Experiences have additionally steered future tasks involving biometric identification methods to restrict bot exercise on potential social platforms.

Filed in Robots. Learn extra about AI (Artificial Intelligence), ChatGPT and OpenAI.