DeepSeek and Moonshot Trained on GPT-6 Astra's Answers

In early September, OpenAI released GPT-6 Astra, its most powerful model at the time. Demand for it was unusually high right away. According to The Information, part of that interest came from Chinese AI researchers who wanted to reverse-engineer Astra's abilities into their own models. Two of China's largest AI companies, DeepSeek and Moonshot AI, reportedly leaned on Astra's answers while developing their latest models.

So Chinese model benchmarks deserve a healthy dose of skepticism. When one model trains on another model's answers, some of its strength actually belongs to that other model. It is better to build new capabilities than to copy ones that already exist.