AI blogger Da7em: China’s models have stalled after Anthropic’s anti-distillation measures
Overseas AI blogger Da7em (X account @Da7_Tech) posted on November 11 local time that after Anthropic strengthened its anti-distillation measures, iteration of China’s large models has clearly slowed. He pointed out that leading labs previously launched significantly stronger models every one or two months, but now there have been no comparable breakthroughs for several months.
Da7em cited several examples: K3.1 Pro, released about 100 days after Kimi K3, did not show a major breakthrough; DeepSeek has been slow to release a Pro-level flagship that satisfies him; Zhipu previously launched a new flagship in less than two months, but now about three months have passed; and Alibaba’s Qwen4-Max preview version has mostly received disappointing feedback. He attributed this to Anthropic’s anti-distillation measures. In February, Anthropic accused DeepSeek, Moonshot (Kimi) and MiniMax of using about 24,000 violating accounts to make more than 16 million calls to Claude in order to extract capabilities; in October, it disclosed that new protections had reduced the reasoning content available to unauthorized users by about 90% to 99%.
Da7em said that Chinese labs previously quickly caught up by distilling large numbers of Claude responses, but after being restricted, their pace has slowed; Kimi and DeepSeek have turned to distilling OpenAI GPT-6 Astra, yet still have not met expectations, which further shows that Claude was previously a key source of capability. He had publicly opposed exaggerating the distillation allegations in August, and in October said Kimi K3.1 Pro was the world’s best open model. Now he has changed his view, saying that Chinese labs have excellent researchers but rely too heavily on others’ capabilities, and need to develop their own technical path rather than merely acting as cheaper substitutes.
Original source:
Title: Strengthening our defenses against distillation attacks
Publisher: Anthropic
Publication date: 2026-10-12
URL: anthropic.com/news/distillation-defenses
Source nature: official company announcement
Verification status: Full text read, confirming the February allegation data, the October statement that 90%-99% less could be extracted, and the description of the protective measures; cross-consistent with reporting by Bloomberg, The Verge and others
Supporting content: Anthropic said it has reduced the reasoning data available to distillers by about 90% to 99%; in February it disclosed that DeepSeek, Moonshot, and MiniMax carried out more than 16 million exchanges through about 24,000 violating accounts
ABAB AI Insight
Historically, transferring frontier model capabilities through output distillation has been a common path in competition between open and closed models. U.S. labs built their advantage early through large amounts of public data and compute investment, while Chinese labs rapidly aligned capabilities at lower cost through API calls, creating a “catch-up window.” When the source company tightens reasoning traces and account controls, that window shrinks.
On the capital path, distillation lowered the pretraining and alignment costs for catch-up players. After Anthropic strengthened its protections, Chinese labs need to turn to their own data, synthetic data, or outputs from other closed models. If OpenAI and others also tighten later, the marginal iteration cost for Chinese labs will rise, and their dependence on overseas compute rental and local cluster scheduling may increase accordingly.
In peer comparison, Anthropic is currently one of the most systematically disclosed vendors of anti-distillation measures. Its “preserved thinking” and two-stage classifiers directly target reasoning extraction. Other U.S. labs have also faced similar allegations, but they have disclosed less about protections and quantified effects. If Chinese labs continue to rely on external outputs, their autonomy in model differentiation and safety alignment will be constrained.
Structurally, anti-distillation shifts capability transfer from being “accessible through public APIs” to requiring bypassing multiple layers of protection. This does not automatically prove that Chinese models as a group have stalled, but it raises the threshold for sustained rapid catch-up. If labs cannot form alternative paths through their own research and data, the iteration pace may shift from “monthly significant leaps” to “longer-cycle incremental improvements,” while more resources are diverted to cluster scheduling and compliant access.
ABAB News · Laws of cognition
Once the reasoning traces of frontier models are locked down by protections, the pace of catch-up that relies on output distillation shifts from monthly leaps to a longer cycle.
The effectiveness of anti-distillation does not depend on a single ban, but on simultaneously restricting proxy networks, context editing, and cross-model reading.
When labs treat “others’ outputs” as their main source of capability, upgrades in the source company’s protections directly translate into iteration costs and differentiation pressure.