Apparently, after Anthropic made it much harder (a.k.a. 90-99% harder) for Chinese labs to distill Claude, the Chinese labs moved on to OpenAI. Per The Information article below, Kimi K3.1 Pro and DeepSeek's V4.1 Flash both lean heavily on GPT-6 Astra instead. By the way, Anthropic apparently told investors that about 5-7% of revenue was due to distillation for the first 9 months of this year. In the prospectus, this is listed as revenue potentially from violating terms of service, etc., but verbally they had said it was distillation supposedly Anyway, Moonshot still claimed to have K3.1 Pro beat Astra, the model it was trained on, on a coding benchmark. But it will be very interesting going forward to see what happens when distillation becomes less effective.