Ah, I see now that you were making a claim narrowly scoped to DeepSeek models specifically. Still, Anthropic has made specific claims about deliberate access to

Ah, I see now that you were making a claim narrowly scoped to DeepSeek models specifically. Still, Anthropic has made specific claims about deliberate access to Claude CoT by DeepSeek (e.g. https://www.anthropic.com/threat-intelligence-report-september-2026#illicit-distillation-sep-26 https://www.anthropic.com/threat-intelligence-report-septemb...) that suggest that they find this information useful even if they do not train directly on it in the way that other labs appear to.