大模型网摘Report: DeepSeek and Moonshot Trained Latest Models on GPT-6 Astra Outputs — What We Know
On October 28, The Information reported that Chinese AI developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's flagship GPT-6 Astra model to train their latest releases, citing people with direct knowledge of the matter; Reuters followed up on October 29. The two training targets named in the report are V4.1-Flash-1017, which went live on October 17, and Kimi K3.1 Pro, launched October 23.
What Distillation Is, and Why It's Sensitive
Distillation means training one model on the outputs of a stronger one, letting the student pick up similar capabilities at far lower cost. For API providers, such use typically violates terms of service: customers paying per token buy inference, not a dataset for training competitors. Since early 2025, questions about whether Chinese labs distill US frontier models have flared up repeatedly, and the US-China AI rivalry amplifies each episode.
Timeline: July to October
Assembling the public reporting: first, OpenAI said on September 30 that it had linked a distillation campaign begun in July to individuals associated with Moonshot; second, The Information reported the two companies turned to Astra after Anthropic hardened its defenses against such scraping, with Anthropic previously claiming its countermeasures cut scraped data by 90-99%; third, GPT-6 Astra, OpenAI's flagship released in early September, has seen strong demand since launch. All of these elements come from media reporting, not company announcements.
Responses: None Public as of This Writing
As of both Reuters' publication and this article, OpenAI, DeepSeek and Moonshot AI had not responded to comment requests. Whether, how, and how much Astra outputs were used remains unconfirmed and undismissed at the corporate level, so readers should treat this as an allegation awaiting verification rather than a finding. For technical background on the two models, see our K3.1 Pro release coverage and first-week benchmark notes on V4.1-Flash-1017.
Three Things to Watch
Technically: if a distillation pipeline is confirmed, the benchmark jumps of both models gain an alternative explanation, and third-party re-testing becomes more meaningful. Commercially: API terms and the enforcement of "no training on outputs" clauses will return to procurement checklists — worth reviewing the official model names and access terms before integrating. Competitively: the open-weight spec race continues, and scrutiny of training-data provenance is likely to intensify; see our October open-weight flagship comparison. This page will be updated if any party responds formally.