Codex With GPT-5.5 Achieves 10× Speedup in NVIDIA Research

NVIDIA has achieved a ten-fold increase in the speed of its end-to-end research workflows through the implementation of Codex, powered by GPT-5.5. Engineers are now utilizing Codex as their primary tool for complex engineering tasks, extending beyond simple code generation to actively identify and resolve issues within programs. “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,” notes Dennis Hannusch, a Senior Software Engineer at NVIDIA. The company’s coding agents team recently leveraged Codex to build a fully functional, internal podcast recording application, comparable to Riverside, in a matter of hours, a task that previously required weeks of procurement and development.

Codex with GPT-5.5 Accelerates NVIDIA’s Research Workflows

NVIDIA’s research and development cycles are experiencing substantial acceleration thanks to the integration of Codex, powered by GPT-5.5, into core workflows. This active debugging capability represents a significant leap forward in AI-assisted development, moving beyond automation to genuine problem-solving. The impact extends to rapid prototyping, as evidenced by the NVIDIA coding agents team’s creation of an internal podcast recording application, functionally similar to Riverside, completed using Codex. Hannusch notes this speed was previously unattainable, citing privacy constraints that would have necessitated weeks of procurement for conventional software solutions.

The team leveraged the Codex desktop app’s computer interaction features to autonomously build and test video and audio functionality, fundamentally altering the scope of feasible projects. “Codex has completely changed the threshold for what’s worth building.” Quantifying the gains, Shaunak Joshi, an AI researcher at NVIDIA, reports that there has been a 10x speed improvement in running experiments, because it’s able to handle the whole end-to-end machine learning research workflow. Joshi also highlights GPT-5.5’s creative potential, noting its ability to synthesize information from research papers and construct knowledge graphs, and its proficiency in code translation, achieving up to a 20x performance increase when converting Python repositories to Rust.

GB200/GB300 Infrastructure Powers Autonomous Codex Sessions

NVIDIA’s shift toward autonomous coding sessions is now underpinned by its GB200 and GB300 infrastructure, enabling a level of sustained performance previously inaccessible to its engineers and researchers. The system operates within this hardware ecosystem. This isn’t merely about accelerating code generation; the system actively participates in debugging, identifying issues beyond the scope of earlier models. “I didn’t have to do anything—it was built and tested completely autonomously,” he states, suggesting a fundamental change in the feasibility of project scope. For research teams, Codex has largely automated the entire research loop, from initial hypothesis to experiment execution.

Codex Enables Rapid Production System Development & Code Translation

Hannusch now relies on the system as his primary tool for tackling complex projects, noting its ability to identify previously undetected flaws. This active debugging represents a substantial advancement, particularly given the previously lengthy timelines associated with similar tasks. “Given our privacy constraints, it would have taken us weeks to procure software,” Hannusch explains, emphasizing the efficiency gains. Many developers are taking their Python repositories, sending them to GPT-5.5, and rewriting them into Rust, resulting in up to a 20x performance increase.