Chinese AI Labs Shift to Distilling OpenAI's Astra Model
nashnova research
Moonshot AI and DeepSeek have redirected their distillation efforts from Anthropic to OpenAI's flagship GPT-6 Astra, after Anthropic's new defenses cut extractable data by 90%–99%. The US-China frontier-model arms race is entering a new phase.
Who is distilling Astra, and what did they get?
Moonshot AI distilled Astra's outputs while developing Kimi K3.1 Pro, released last week, according to three people familiar with the matter.
DeepSeek's V4.1-Flash 1017, released October 17, also used Astra distillation data. The update showed significant improvement over the September original.
In plain terms = distillation — using a stronger model's outputs as "textbook material" to train your own smaller model — is not stealing code. It means making massive API calls and harvesting the answers as training data.
Why did the target shift from Anthropic to OpenAI?
Anthropic's Claude series was previously the top distillation target for Chinese labs. The shift happened because Anthropic announced "material progress" in fighting distillation this month.
New defenses include restricting proxy-network access and imposing stricter rate limits and bans on users who repeatedly try to extract hidden reasoning chains — cutting extractable data by roughly 90% to 99%.
This means → Anthropic nearly shut off the "textbook supply," so distillers naturally pivoted to OpenAI's less-defended Astra.
How much revenue does distillation generate for Anthropic and OpenAI?
Anthropic's IPO prospectus disclosed that customers violating its terms of service contributed roughly 5%–7% of revenue in the first nine months of this year. One Chinese proxy-network operator estimated its clients paid Anthropic about $10 million per month.
Anthropic's total revenue for the first nine months exceeded $33 billion. This reflects that distillation traffic has become a non-trivial revenue stream.
On the OpenAI side, annualized revenue reached nearly $50 billion last month, up roughly 70% from end-June — but how much distillation contributed remains unclear.
How is OpenAI responding?
An OpenAI spokesperson said the company takes distillation "extremely seriously" and is actively advancing countermeasures, but declined to estimate distillation's revenue impact.
The spokesperson's words: "As we deploy new defenses, attackers adjust their tactics — that requires us to keep improving."
This means → OpenAI acknowledges a dynamic cat-and-mouse game, but has not yet delivered a concrete result matching Anthropic's "90% compression" benchmark.
Is distillation the only way Chinese AI closes the gap?
No. US export controls limit Chinese labs' access to advanced chips (GPUs), which has pushed DeepSeek and others to develop innovative techniques for training and running models with less compute.
But Chinese lab researchers stated plainly: distillation has played a critical role in helping Chinese AI stay close to frontier models at far lower compute costs than US peers.
In plain terms = distillation is a shortcut — skip training from scratch and use a strong model's answers as your textbook, saving massive compute and time. Whether OpenAI can block this shortcut by 90% or more, as Anthropic did, is the next key test in this arms race.
市场有风险，内容仅供研究参考，不构成投资建议。