Chinese AI models claim 30%+ of U.S. token traffic; Z.ai's GLM-5.2 rivals Anthropic on cost and code
U.S. companies are rapidly shifting to Chinese-built AI models as costs for frontier systems from Anthropic and OpenAI surge. According to OpenRouter, a platform aggregating multiple AI model APIs, Chinese models account for over 30 percent of tokens consumed each week by U.S. developers, up from just 11 percent average in the prior 12 months and 4.5 percent in H1 2025. The shift reflects cost-conscious teams routing inference to cheaper alternatives when performance permits, driven by token price increases at major U.S. labs.
Beijing-based Z.ai's GLM-5.2 model, launched in June, has catalyzed a 'mini DeepSeek moment' among developers. Industry observers describe GLM-5.2 as operating at roughly one-sixth the cost of Anthropic's Claude Opus 4.8 and OpenAI's GPT offerings, while delivering comparable coding and agentic AI capabilities—within a percentage point of Opus 4.8 on one closely watched benchmark. Z.ai claims the 1 million token context window enables handling of long-running engineering tasks. On Vercel's platform, GLM-5.2 saw the fastest adoption of any model in 2026, with daily token volume growing 27x in its first week.
For practitioners, the cost arbitrage is now material: Chinese open-source models cost 60-90% less than frontier U.S. offerings for overlapping workloads. However, adoption remains segmented by geography and regulation. U.S. enterprises in banking and cybersecurity cite data security concerns as barriers. Startups and SMEs are adopting faster. Regulatory uncertainty—restrictions on Anthropic's latest models and delayed public rollout of OpenAI's GPT-5.6—is accelerating the trend. The architectural implication: teams building systems that can route work across multiple model APIs now have a cost advantage, and multi-model inference patterns will become standard.