China's Top AI Labs Reportedly Distilled GPT-6 Astra to Build New Models
Moonshot and DeepSeek reportedly used OpenAI's GPT-6 Astra outputs to train their latest models, raising questions about their true edge.
Models · Source: Reuters
What happened
Two of China's top AI labs reportedly took a massive shortcut to catch up with American competitors. Moonshot AI and DeepSeek are accused of using outputs from OpenAI's GPT-6 Astra to train their newest models. This technique is known as distillation. It works like a junior chef tasting a master's dishes to steal the recipe. Engineers feed a powerful model thousands of prompts, collect the high-quality answers, and use those responses to train a different, cheaper model.
DeepSeek launched DeepSeek-V4.1-Flash-1017 on October 17. The update focuses heavily on sharper task performance and better agentic coding. Agentic coding means the model does not just suggest snippets of code. It actively plans and carries out multi-step programming work with minimal human hand-holding. DeepSeek kept its signature approach intact, keeping the model open-weight and low-cost. Moonshot followed on October 23 with Kimi K3.1 Pro. This massive upgrade packs approximately 2.8 trillion parameters. Parameters act as the internal dials a model tunes during training, giving it a massive capacity to learn complex patterns.
The irony is that the student is now outscoring the teacher on specific tests. Kimi K3.1 Pro scored a 55 on the Artificial Analysis Intelligence Index, placing it fourth overall. It even beat GPT-6 Astra on the Terminal-Bench 4.0 agentic benchmark. However, this is not the first time these labs have faced these allegations. Back in February, Anthropic accused both Moonshot and DeepSeek of running distillation attacks on its Claude models. Anthropic claimed the labs used fraudulent accounts to extract the outputs.
Key facts
- 2.8 trillion — Parameters in Moonshot's new Kimi K3.1 Pro model
- 55 — Kimi K3.1 Pro's score on the Artificial Analysis Intelligence Index
- $50 billion — Reported valuation of Moonshot AI ahead of its planned IPO
Why it matters
If you are building AI products, this changes the math on competitive moats and training costs. Training frontier models from scratch requires billions of dollars and massive data centers. Distillation proves you can achieve state-of-the-art performance at a fraction of the cost by simply scraping a competitor's API. It turns a hardware and research race into a data extraction game. Both labs are pitching their models for practical jobs like financial analysis and coding, directly threatening the market share of closed US systems.
The second-order effect hits the venture capital markets and international policy. These are not scrappy side projects. Moonshot AI reached a reported valuation of around 50 billion dollars ahead of a planned IPO in Hong Kong. DeepSeek has been valued even higher in recent funding rounds. That is a massive amount of capital riding on the assumption that Chinese labs can compete at the frontier. If their core technical advantage relies on OpenAI's outputs, that valuation looks fragile. Investors will start demanding proof of proprietary training data. Furthermore, if OpenAI decides to publicly retaliate, the debate over who owns a model's outputs could quickly escalate into formal policy and trade restrictions.
For builders
Leverage cheap distilled models for agentic workflows
DeepSeek and Moonshot are proving that distilled open-weight models can match or beat frontier models on specific coding tasks. Builders can swap expensive OpenAI API calls for these cheaper alternatives to run multi-step programming agents. You save money while getting top-tier performance for specialized tasks.
Expect strict API crackdowns from frontier labs
OpenAI and Anthropic lose their competitive edge when rivals scrape their outputs. Expect them to deploy aggressive rate limits, account bans, and output watermarking to stop distillation attacks. If your startup relies on generating synthetic training data from GPT-6 Astra, your pipeline might break without warning.
Scrutinize the moats of open-weight providers
The low-cost, open-weight playbook is highly attractive to developers. But if questions about where a lab gets its training signal start to stick, their cost advantage may vanish under legal or technical pressure. Do not build your entire infrastructure around a single open-weight model that might face IP lawsuits.
My take
Building a 50 billion dollar company by copying OpenAI's homework is a bold move, but it is not a durable moat. Distillation is a brilliant shortcut to catch up. However, you cannot lead the market if you rely on your rival's API for your core training signal. The real winners right now are the developers getting frontier performance at open-weight prices.
Original reporting: Reuters. This is my rewrite and opinion.