Chinese AI development labs, including DeepSeek and Moonshot AI, have shifted their model distillation efforts to focus primarily on OpenAI’s GPT-6 Astra. The labs are extracting output reasoning steps to train recent models like Kimi K3.1 Pro and V4.1-Flash 1017, according to a report from The Information.

This shift occurred after Anthropic implemented stronger defensive measures on its Claude models, which reduced potential data extraction rates by 90 to 99 percent. Distillation relies on forcing frontier models to expose hidden chain-of-thought outputs, a practice explicitly prohibited by both OpenAI and Anthropic terms of service.

OpenAI previously disclosed a coordinated distillation campaign linked to external actors in late September. A company spokesperson stated OpenAI considers distillation a severe threat and continues to update defenses as rival labs adapt their scraping methodologies.

Why it matters

  • Demonstrates the increasing role of model distillation in bridging the gap between frontier models and open-weight alternatives.

  • Highlights the technical arms race around chain-of-thought privacy and API extraction defenses.

  • Signals ongoing geopolitical tension around frontier AI intellectual property and training data source controls.

Source: the-decoder.com