OpenAI is ramping up restrictions to prevent artificial intelligence developers in China from harvesting the reasoning capabilities of its GPT-6 Astra model. The move aims to stop the practice of distillation, where labs use a powerful “teacher” model to generate responses that are then used to train smaller, cheaper “student” models, bypassing the immense cost of building frontier systems from scratch.
The Information reports that Chinese firms DeepSeek and Moonshot recently used Astra to develop their V4.1-Flash 1017 and Kimi K3.1 Pro models after a similar crackdown at Anthropic. Anthropic stated that its use of account bans and proxy blocks reduced the data harvested from its Claude model by 90% to 99% in recent months. While OpenAI traced a July reasoning-extraction campaign to individuals linked to Moonshot, Chinese developers have argued that U.S. GPU export restrictions necessitate such training methods.
