Chinese AI companies DeepSeek and Moonshot AI relied on OpenAI models to train their latest software, according to The Next Web reporting on 29 October 2026. The information comes via The Information, which cited people with direct knowledge of the situation involving the use of GPT-6 Astra.
Key takeaways
- DeepSeek and Moonshot trained new models using GPT-6 Astra data.
- The process involves distillation to replicate advanced capabilities cheaply.
- OpenAI previously accused Moonshot of running a data extraction campaign.
- Anthropic recently introduced safeguards to block similar data harvesting.
How do Chinese labs use OpenAI models for training?
The method is called distillation, and it lets developers learn from a stronger system instead of building features from scratch. Teams send thousands of prompts, gather the output, and train their own software on those replies. Done at scale, this lets smaller systems match advanced abilities at a lower price.
The specific software includes Moonshot's Kimi K3.1 Pro, which launched last week, and an updated DeepSeek V4.1-Flash version that debuted on 17 October. Both systems posted strong benchmark results, though those scores remain unverified independently.
What does the AI industry say about data distillation?
OpenAI previously stated that individuals connected to Moonshot ran a coordinated campaign in July to pull hidden reasoning data from its systems. OpenAI blocked the accounts involved by 28 July after detecting activity across more than 15,000 users.
US labs like OpenAI, Anthropic, and Google argue that this practice lets rivals copy advanced features without spending comparable amounts on research and training. US agencies have also pointed to Chinese firms targeting American systems.
Why it matters
Leading AI labs face a tough challenge because the same tools they provide to everyday users can double as training data for rival firms. To counter this, Anthropic implemented new defenses earlier this month that cut data harvesting from Claude models by 90% to 99%, showing how major developers intend to protect their work.
Frequently asked questions
DeepSeek and Moonshot AI relied on OpenAI models, specifically GPT-6 Astra, to train their latest software according to a report cited by The Next Web.
Distillation is a process where developers send large numbers of prompts to a capable model, collect the answers, and use those responses to train their own systems at a fraction of the cost.
Moonshot released its Kimi K3.1 Pro last week, while DeepSeek launched an updated version of its V4.1-Flash model on 17 October.
OpenAI stated that people linked to Moonshot ran a coordinated campaign in July involving over 15,000 users to extract hidden reasoning data from its models.
Anthropic introduced new safeguards earlier this month that reduced data collection from its Claude models during distillation attacks by around 90% to 99%.
Based on reporting by The Next Web. Read the original article
