tl;dr: The author reports that Kimi K3 matches Claude's coding output quality and token efficiency at roughly a third of the price, while Claude's subscription plans quietly downgrade users off headline models due to economics. Open Chinese models like Kimi K3 and GLM 5.2 (MIT-licensed) are also outperforming restricted US models on benchmarks like Semgrep's cyber tests because they don't refuse work. The author argues US AI policy is backfiring—hindering domestic models while doing nothing to constrain freely-downloadable foreign alternatives—and predicts protectionist measures will leave American users stuck with inferior, more expensive options.