Sources: DeepSeek and Moonshot AI heavily relied on GPT-6 Astra distillation to train models like Kimi K3.1 Pro and V4.1-Flash 1017, after Anthropic's crackdown

The core capability of many advanced AI models becomes a readily available commodity for competitors, lowering barriers to entry.

Chinese AI developers DeepSeek and Moonshot AI reportedly used OpenAI's GPT-6 Astra model for knowledge distillation to train their latest models, including Moonshot AI's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017. This practice, where a smaller model learns from a larger one, has become contentious amid US-China tech competition, with companies like OpenAI and Anthropic arguing it allows rivals to replicate advanced AI capabilities rapidly and cost-effectively. OpenAI had previously indicated it linked Moonshot AI to a distillation campaign. Anthropic recently announced measures that reduced data harvested through such attacks by 90% to 99%, suggesting a broader effort by major AI labs to curb this practice.

The reliance on GPT-6 Astra distillation by Chinese firms highlights a critical vulnerability in the current AI development landscape: the ease with which proprietary model capabilities can be replicated. This practice circumvents the immense cost and time typically required for foundational model training, potentially accelerating the pace of innovation and competition, but also raising concerns about intellectual property and fair play in the global AI race.

As major AI labs like OpenAI and Anthropic implement stronger defenses against distillation, developers who previously relied on this method will need to pivot to alternative strategies for model improvement. This could involve increased investment in proprietary data acquisition, novel training methodologies, or exploring open-source ecosystems, signaling a potential shift in the economic dynamics and competitive strategies within the AI development sector.