China's AI labs shift distillation target from Claude to OpenAI's GPT-6 Astra, report says

What's the deal? Chinese AI firms DeepSeek and Moonshot AI reportedly trained new models using outputs from OpenAI's GPT-6 Astra, according to The Information, which cited unnamed sources. Astra, released in September, has become a primary target for so-called distillation.

What's distillation? It is a technique that transfers a stronger model's outputs into the training of another, cutting the cost and time of building a competitor. Anthropic and OpenAI use it themselves, but both ban outside firms from distilling their models and do not make them available in China.

The details: Sources said Moonshot distilled Astra while developing "Kimi K3.1 Pro," released last week, and DeepSeek drew on Astra for its "V4.1-Flash 1017" version, released on October 17. Neither company has confirmed the claims.

Why the switch? Chinese labs had until recently focused on Anthropic's Claude, but sources said they shifted to OpenAI after Anthropic took strong countermeasures. Anthropic said this month its new safeguards left distillers with 90% to 99% less data to harvest.

The wider picture: A September 8 advisory from the National Security Agency, Federal Bureau of Investigation, and Cybersecurity and Infrastructure Security Agency named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI, alleging they targeted Claude, GPT, Gemini, and Grok since at least late 2024. China's Ministry of Commerce called the claims "groundless and legally unsound."

What could go wrong? The reports remain unverified. Headlines retain "reportedly" because the alleged training use has not been confirmed, and the claims do not prove the Chinese models are mere copies.

The signal: The change in target suggests distillation is not a Claude-specific problem but a wider contest over whether frontier labs can protect the capabilities they sell through APIs.

Image credit: Generated with Gemini

Read more: thenextweb.com, The Information