GPT-5.5-powered Codex Boosts Nvidia Employees' Efficiency, Cutting Cost per Million Tokens by 35x

Nvidia recently announced that its internal tool Codex is powered by OpenAI's latest model GPT-5.5 and runs on the Nvidia GB200 NVL72 rack-level system. Currently, Codex has been adopted by multiple teams including engineering, legal, and marketing, with over 10,000 employees giving feedback that the effect is 'shocking'.

The system brings significant efficiency improvements: compared to the previous generation, the GB200 NVL72 reduces the cost per million tokens by 35x and increases the token output per megawatt per second by 50x. Debugging that used to take days now takes only hours, complex code experiments are shortened from weeks to overnight; teams can also deliver end-to-end functions via natural language prompts, with higher reliability and less rework.