In 2026, Chinese labs moved from one breakout to an ecosystem: DeepSeek V4, V4-Flash, Moonshot Kimi, http://Z.ai GLM-5.2 - five frontier-level models in eight weeks. The wedge is price and efficiency.
Independent testing shows a complex workload costs $0.03 with DeepSeek V4 Flash vs $3.15 with Claude Fable 5. OpenAI flagship GPT-6 Astra is $10/$50 per million tokens in/out, DeepSeek V4.1-Flash is $0.035/$0.12 - roughly 300-400x cheaper. c554dd04
That forces US labs to cut: OpenAI slashed 80% on its Luna model, Anthropic cut price in half. eaf1
Threat is 3-fold:
1. Market share: Open-weight models in production jumped from 23% to 67% in one year. Global South buyers don't want to pick sides.
2. Stock shock: January 2025 DeepSeek launch wiped $1T, including Nvidia's $600B record loss.
3. Security: DHS banned DeepSeek from civilian federal systems in Oct 2026, warning it "presents multiple threat vectors" and risks data transfer to China. 75714fc5
So threat is very big on pricing, moderate on absolute frontier performance - but closing fast.