How NVIDIA uses OpenAI Codex with GPT-5.5

NVIDIA integrates Codex and GPT-5.5 for autonomous engineering

NVIDIA engineers and researchers have adopted Codex, powered by GPT-5.5 and running on NVIDIA GB200 and GB300 infrastructure, as their primary tool for complex engineering and machine learning research. The system enables longer, more autonomous sessions that can identify bugs, gaps, and new ideas beyond the initial prompt.

Production system development and deployment

Codex with GPT-5.5 allows NVIDIA engineers to move projects from minimum viable product (MVP) to production-ready systems with improved scalability and reliability. The tool's autonomy reduces the need for manual guidance during long development sessions while maintaining high accuracy and context.

Key applications in production engineering include:

• Rapid Application Development: NVIDIA engineers used Codex to build an internal podcast recording application in hours, bypassing weeks of software procurement processes due to privacy constraints.
• Autonomous Testing: Using the Codex desktop app with computer interaction capabilities, the system autonomously built and tested video and audio recording functionality without manual intervention.
• Tool Selection: The model demonstrates a high capacity for tactically selecting the appropriate tools and skills for specific engineering tasks.

"Codex is our go-to tool for complex engineering tasks, and with GPT-5.5, it surfaces bugs and gaps in my program that other models weren’t able to find."

—Dennis Hannusch, Senior Software Engineer

Automation of machine learning research workflows

Codex has automated the end-to-end research loop for NVIDIA's AI researchers, including identifying research areas, writing experiment scripts, and executing those scripts on remote machines. This has resulted in a reported 10x speed improvement in running experiments.

Technical capabilities utilized in research include:

• Knowledge Synthesis: By analyzing large corpora of research papers in fields like reinforcement learning, GPT-5.5 acts as a creative partner, tracing evidence across chains and suggesting knowledge graphs to visualize conceptual connections.
• Remote Infrastructure Management: Because the Codex app supports SSH, researchers can run large machine learning workloads on remote hosts directly from their laptops without manual login and setup.
• Performance Optimization: Researchers are using GPT-5.5 for "machine translation" of codebases, specifically rewriting Python repositories into Rust to achieve up to 20x efficiency gains.

"It’s been a 10x speed improvement just in terms of running experiments, because it’s able to handle the whole end-to-end machine learning research workflow."

—Shaunak Joshi, AI researcher