DeepSeek and Moonshot used OpenAI's GPT-6 Astra to train latest models, report says
Chinese AI labs DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to build their newest models after an Anthropic crackdown restricted access to Claude, according to The Information.
Chinese artificial intelligence developers Moonshot AI and DeepSeek relied heavily on outputs from OpenAI's GPT-6 Astra to train their newest models, according to an exclusive report by The Information. Moonshot reportedly distilled Astra while developing Kimi K3.1 Pro, released in late October, while DeepSeek's Oct. 17 update, V4.1-Flash 1017, also benefited from the technique and showed substantial gains over its September predecessor. The reporting cited three people with direct knowledge of Moonshot's development and two familiar with DeepSeek's work, though neither lab has commented publicly on the claims.[1][2][3][4]
The pivot to OpenAI's flagship model followed strict counter-measures by Anthropic, whose Claude models previously served as the primary target for Chinese distillation campaigns. Distillation allows developers to train smaller models by querying a more powerful system through standard APIs rather than training frontier architectures from scratch. In February, Anthropic revealed that DeepSeek, Moonshot, and MiniMax used roughly 24,000 fraudulent accounts to log more than 16 million chats with Claude. Anthropic subsequently began hiding Claude's step-by-step reasoning outputs, reporting that its defenses cut data harvesting by 90% to 99% and prompting labs to target Astra following its early September release.[1][3][4][5][7][8][9]
DeepSeek founder Liang Wenfeng initially intended to avoid distillation while developing the V4 model generation, but reversed course due to competitive pressure after Chinese rivals achieved traction using the method, according to the report. Distillation efforts against OpenAI had already triggered scrutiny before the latest releases, with OpenAI stating in September that it had traced a campaign launched in July directly to individuals linked to Moonshot.[2][4][5]
Key facts
- The Information reported that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train their latest models.
- Moonshot AI used Astra outputs to build Kimi K3.1 Pro, while DeepSeek's Oct. 17 update, V4.1-Flash 1017, also distilled the model.
- The reporting is based on three people with direct knowledge of Moonshot's work and two familiar with DeepSeek's; neither lab has commented.
- The shift toward Astra occurred after Anthropic deployed safeguards on Claude that cut distillers' harvested data by 90% to 99%.
- Anthropic previously disclosed in February that DeepSeek, Moonshot, and MiniMax used approximately 24,000 accounts to run over 16 million chats with Claude.
- OpenAI said in September that it had traced a distillation effort originating in July to individuals connected to Moonshot.
- DeepSeek founder Liang Wenfeng initially sought to avoid distillation for the V4 generation but altered strategy under domestic market pressure.
Sources · 7 sources
- TI
The Information@theinformationPost on X ·
Exclusive: DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train their latest models. The shift comes as Anthropic cracks down on Chinese labs extracting capabilities from Claude. Full story: https://t.co/knT5nXLx2H
Open source - WS
Wall St Engine@wallstenginePost on X ·
DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train their latest models, The Information reports. Moonshot used Astra while developing Kimi K3.1 Pro, released last week. DeepSeek’s October 17 update, V4.1-Flash 1017, also benefited from distilling Astra and showed substantial improvements over its September version. Astra itself only launched in early September, putting those releases within weeks of the model they reportedly drew from. The report also says DeepSeek’s Liang Wenfeng initially wanted to avoid distillation when developing the V4 generation, but changed course under competitive pressure after seeing rivals such as Z .ai gain traction with GLM models that also used the technique. The reporting cites three people with direct knowledge of Moonshot’s work and two familiar with DeepSeek’s.
Open source - CH
Chubby♨️@kimmonismusPost on X ·
As expected, Kimi 3.1 Pro distilled US frontier models to train their models. This time they distilled GPT-6 Astra. The damage to Kimi’s reputation keeps growing. Via the information https://t.co/loSXL7Ruft
Open source - MT
MTS@MTSlivePost on X ·
SITUATION EXPLAINED: Anthropic cut off Chinese labs from distilling Claude, so they reportedly switched to OpenAI's GPT-6 Astra. • Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017, both released this month, were reportedly trained on Astra's outputs, and neither lab has commented • Distillation means training a smaller "student" model to copy a stronger "teacher" model's answers, and it works through a normal API without the teacher's weights • In February, Anthropic said DeepSeek, Moonshot and MiniMax used about 24,000 fake accounts to run over 16 million chats with Claude • Distillers mostly want Claude's step-by-step reasoning, and in Anthropic's own test, a model trained on it scored 80.4% on a coding benchmark vs 71.8% when trained on summaries • Anthropic now hides that reasoning from distillers and bans their accounts, often within 2 hours of a campaign starting • Anthropic reportedly told investors distillation made up 5% to 7% of its revenue, and it's walking away from that money to shut it down • OpenAI said in September that it traced a distillation campaign that began in July to people linked to Moonshot @theojaffee: "It's not all distillation. I think there's a little bit of cope going on here. They genuinely have amazing talent. They're genuinely able to work very well with very limited resources compared to what US labs have in terms of compute."
Open source - MN
Mario Nawfal@MarioNawfalPost on X ·
🇨🇳🇺🇸 China's hottest new AI models reportedly have a secret ingredient: OpenAI's homework. DeepSeek and Moonshot AI leaned heavily on outputs from OpenAI's powerful GPT-6 Astra to train their latest releases, including Moonshot's Kimi K3.1 Pro and DeepSeek's upgraded V4.1-Flash, both out this month. The trick is called distillation. You bombard a top AI with questions, collect its answers, then use them to train your own model. A shortcut to advanced AI capabilities without the eye-watering development bill. Why spend billions building a frontier model when you can just ask one nicely? Sources: Reuters, The Information / Writer: Julie
Open source - TI
The Information@theinformationPost on X ·
Exclusive: OpenAI’s most powerful model has become a major target for Chinese AI labs seeking to improve their own models. DeepSeek and Moonshot used Astra outputs to help train recent releases, according to people with direct knowledge. Full story: https://t.co/knT5nXLx2H
Open source - AG
AGTP@AGTPinsightsPost on X ·
Chinese AI labs DeepSeek and Moonshot AI just turned to OpenAI's GPT-6 Astra as a distillation target. Here's what you need to know. The Information reports both labs drew on Astra's outputs to train new models released this month. That includes Moonshot's Kimi K3.1 Pro. Reuters says they relied heavily on those outputs. Distillation means training a model on another model's answers. The shift follows Anthropic's crackdown. In February, Anthropic said it found industrial-scale distillation campaigns by DeepSeek, Moonshot and MiniMax, using about 24,000 fraudulent accounts to extract Claude's capabilities. In October it added new defenses, including stricter usage limits for users who keep trying to extract reasoning. Anthropic says those cut distillers' harvested data by 90-99%. With Claude harder to harvest, Astra has become the main target. Astra is OpenAI's newest model. OpenAI reports 57.9% on one benchmark, against 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1. It also reports 1.9x faster task completion than GPT-5.6 Sol on Mind2Web. The timing hurts Moonshot. Artificial Analysis released Intelligence Index v5 on Oct 28, with the private share of its benchmarks raised from 45% to 60%. Kimi K3.1 Pro's score dropped from 55 to 46. The Information's report landed a couple of hours later. Analysts say the models may now share Astra's style, and some argue a lab that distills can't count as frontier. Key numbers: - Astra: 57.9% on the cited benchmark vs 37.3% for GPT-5.6 Sol - Claude Fable 5.1: 55.8% on the same benchmark - Kimi K3.1 Pro on Artificial Analysis: 55 to 46 - Private share of the AA index: 45% to 60% - Anthropic's defenses: 90-99% less harvested data - About 24,000 fraudulent accounts in the February Claude campaigns The report is based on The Information's findings, and neither lab has been quoted in the material.
Open source - TE
TechmemeArticle ·
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 Information) The Information : 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 — OpenAI's GPT-6 Astra, the most powerful model the company has released, has met with strong demand since its release in early September.
Open source - TI
The Information@theinformationPost on X ·
Exclusive: Anthropic’s crackdown on distillation has pushed Chinese AI labs toward OpenAI’s GPT-6 Astra, now a major source of training data for DeepSeek and Moonshot. Full story: https://t.co/knT5nXLx2H
Open source