Nvidia employees and Huang himself evaluated the performance of OpenAI Codex built on GPT‑5.5.

New Codex Capabilities from OpenAI for Nvidia

OpenAI and Nvidia have launched a local version of the Codex service built on the GPT‑5.5 model. The results of this system impressed employees at the “green” company: efficiency has increased compared to previous generations.

MetricGPT‑4oGPT‑5.5 (Codex)Economic Cost35× higher35× lowerToken output per megawatt–50× higher
The service is available to 10,000 Nvidia employees and has already shown “mind-blowing” results that can change the workflow.

> “Debugging cycles that used to take several days now finish in a couple of hours. Experiments that required weeks are executed instantly even in large multi-file repositories. Teams develop full function packages by sending natural language requests and receive higher reliability with lower resource losses,” – Nvidia representatives said.

Codex was deployed on Nvidia GB200 NVL72 workstation systems, which provided significant cost savings and optimal token consumption.

Nvidia CEO Jansen Huang wrote to Sam Altman (CEO of OpenAI) expressing gratitude for GPT‑5.5’s contribution to speeding up employee work:

> “I am incredibly excited that Codex will help teams accomplish tasks that once seemed impossible. Congratulations to everyone on opening new horizons, and thank you for creating GPT, which has become a springboard for reasoning, planning, and tool usage.”