DeepSeek’s claimed cost advantage and the debate over frontier AI labs

In a September 15, 2026 exchange, a user estimated DeepSeek v4.1 Flash cost $10 million to train—100x less than GPT-6 Astra—and claimed it also costs 100x less to run while scoring about the same on everyday design and other benchmarks. The user asked what the bull case for frontier AI labs would be once open-source models are competent enough.

The disagreement centered on distillation, or learning from other models. A respondent argued that Chinese labs still rely on American models, particularly their thinking traces and ways of breaking down problems, while crediting Chinese video models as frontier-level. The original poster argued that Chinese labs already had enough data to avoid further distillation. The respondent said China could develop models independently but claimed it had not yet demonstrated a clean-slate model.