The performance gap between US and Chinese AI models has collapsed from over 1,300 points in May 2023 to just 2.7% (39 points) by March 2026, according to Stanford’s 2026 AI Index. DeepSeek has been central to this convergence: its R1 reasoning model launched in January 2025, followed by the V4 series in April 2026, putting Chinese AI on the global map with significantly lower costs than US alternatives (DeepSeek V4-Pro charging approximately $3.96 per million output tokens). By mid-2026, Chinese models already captured 50-67% of token processing on OpenRouter. The US still maintains a 6-8 month lead on specialized benchmarks in complex reasoning and cyber tasks, and holds approximately 74% of global compute power.
Source: Read the original article
