According to a report by The Information, leading Chinese artificial intelligence developers DeepSeek and Moonshot AI have increasingly relied on OpenAI’s GPT-6 Astra model for model distillation. By harvesting reasoning steps and outputs from Astra, the labs reportedly trained recent models like Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017. This shift occurred after Anthropic successfully fortified its systems against data scraping, significantly cutting the data available for distillation from its Claude models. While OpenAI and other Western frontier labs strictly prohibit the practice, the trend highlights the ongoing global race to acquire advanced AI capabilities at a fraction of standard training costs.

Full report : DeepSeek and Moonshot AI targeted OpenAI’s GPT-6 Astra for AI training after Anthropic tightened access to its frontier model.