ABAB News Logo
Back to Flash News

DeepSeek and Moonshot Rely on OpenAI GPT-6 Astra Outputs to Train Latest Models

The Information reported on October 28 that Chinese AI developers DeepSeek and Moonshot AI relied heavily on the output of OpenAI’s GPT-6 Astra model when training their respective latest models, the publication said, citing people familiar with the matter. Reuters then relayed the news.

This practice is known as distillation. The Information said that Moonshot AI’s Kimi K3.1 Pro, released last week, as well as DeepSeek’s upgraded V4.1-Flash 1017 launched on October 17, both benefited from distillation of Astra outputs. OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters’ requests for comment.

OpenAI previously said in an official blog post on September 30 that it had linked the core part of a distillation operation that began in July to individuals associated with Moonshot AI. The operation involved attempts to extract protected reasoning content, and OpenAI said it had fully disrupted the related activity by July 28.

Original sources:

  1. Title: OpenAI’s Astra Model Becomes Top Distillation Target for China’s AI Labs Publisher: Jing Yang, Juro Osawa, Qianer Liu (The Information) Publication date: 2026-10-28 URL: https://www.theinformation.com/features/exclusive（文章摘要页面；全文可能需订阅） Source type: media report Verification: On this page read, only the visible summary, title, authors and publication time were accessed; the full text behind the paywall was not accessed Supporting content: DeepSeek and Moonshot AI relied heavily on Astra outputs when training their latest models; the relevant models were Kimi K3.1 Pro and V4.1-Flash 1017

  2. Title: DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train latest models, The Information reports Publisher: Reuters Publication date: 2026-10-28 URL: https://www.reuters.com/technology/deepseek-moonshot-used-openais-gpt-6-astra-train-latest-models-information-2026-10-29/ Source type: media report Verification: The core content was confirmed this time through search snippets and multiple reposts; direct page access was blocked, but multiple reposts were consistent Supporting content: The Information report, distillation practice, the two specific models, companies did not respond, OpenAI’s September 30 blog linked individuals associated with Moonshot

  3. Title: Disrupting a coordinated model-distillation campaign Publisher: OpenAI Publication date: 2026-09-30 URL: https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/ Source type: official announcement Verification: The full page text was read in this visit Supporting content: Activity was observed starting on July 1, a high-volume peak appeared on July 24-25 (16,000 related extraction-pattern requests from more than 4,000 users), a cluster of more than 15,000 users was linked, and disruption occurred by July 28; the core cluster was attributed to individuals associated with Moonshot AI; encryption or databases were not broken

  4. Title: GPT-6 Astra: A new generation of intelligence Publisher: OpenAI Publication date: 2026-09 (page updated with late-September information) URL: https://openai.com/index/gpt-6-astra/ Source type: official announcement Verification: Introductory sections of the page were read in this visit Supporting content: GPT-6 Astra is a model launched by OpenAI, positioned as a frontier model for computer use, software engineering and other capabilities; it does not prove the specific distillation volume or effect in this case

  5. Title: Strengthening our defenses against distillation attacks Publisher: Anthropic Publication date: 2026-10-12 URL: https://www.anthropic.com/news/distillation-defenses Source type: official announcement Verification: Key passages on the page were read in this visit Supporting content: Anthropic said new measures reduced the amount of data distillers could obtain from Claude by about 90% to 99% (company-reported, based on extraction volume of inference tokens on its own platform)

ABAB AI Insight

OpenAI’s blog post published on September 30 attributed the core cluster of the July distillation activity to individuals associated with Moonshot AI, the developer of Kimi. The activity began at low volume on July 1, then on July 24-25 there was a peak of 16,000 related requests from more than 4,000 users, linked to a cluster of more than 15,000 users, and the company said it had disrupted the activity by July 28. This statement concerns attempts to extract protected reasoning content, and does not directly prove the full data composition used in subsequent model training.

The distillation path is this: by querying a stronger model in large volumes to obtain its outputs or reasoning traces, those outputs are then used as training data for one’s own model, allowing some capabilities to be replicated at lower compute and data cost. OpenAI had previously associated DeepSeek and others with similar practices; on October 12, Anthropic said its new measures against Claude reduced related extraction by about 90% to 99%. DeepSeek’s V4.1-Flash 1017 is an open-weight model, while Moonshot’s Kimi K3.1 Pro showed strong agentic and coding performance in independent evaluations, and both were released within weeks after Astra’s launch.

Compared with the case in early 2025 in which OpenAI and Microsoft investigated whether DeepSeek might have made heavy API calls and distilled GPT series models, this report targets GPT-6 Astra, a newer model with computer-use and agentic capabilities, and The Information’s report extends the scope to Moonshot’s latest release. The key similarity is using the other party’s output to accelerate one’s own training; the difference is that OpenAI has publicly disrupted a specific operation and named the associated individuals, whereas the current report still relies on anonymous sources and the companies have not publicly responded.

If this report is accurate, what changes is the cost path for DeepSeek and Moonshot to obtain frontier agentic and coding capabilities in their latest iterations: they could use Astra’s outputs to improve performance on relevant tasks without fully bearing the pretraining cost at Astra’s scale. The observation metrics are subsequent changes in relative scores on independent agent benchmarks such as Terminal-Bench, and whether OpenAI or Anthropic further restricts related access or discloses more evidence.

ABAB News · Cognitive Law

  1. Distillation treats the other side’s outputs as training data, and the compute threshold for capability replication is determined by the other side’s query volume rather than one’s own pretraining scale.

  2. Even after OpenAI disrupted one operation, the latest models of the related companies were still reported to rely on its outputs; public disruption does not mean the downstream data flow is cut off.