A latest version of an unreleased OpenAI model achieved a score of 0.758 on its most basic reasoning level during BenchCAD 2.0 hardware design tests, surpassing the 0.594 maximum score recorded by GPT-6 Astra. The new internal model’s peak performance in the same benchmark reached 0.875, indicating that the new model’s lowest configuration outperforms the most expensive version of its predecessor.
The efficiency gains follow a similar pattern to the company’s recent results on the Navier-Stokes physics problem. In addition to these benchmark scores, OpenAI stated that GPT-6 Astra was able to design a physical gadget’s software, circuit board, and case in less than a week with only two staffers and no direct human edits to the design process. The company is testing these electrical and mechanical engineering workflows to speed up iteration for its internal hardware teams and industry partners.
