待翻译：Chinese AI Models Overtake American Rivals in Popularity

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

Take that, OpenAI! Anthropic! Chinese AI models have surpassed their U.S. counterparts in token consumption on OpenRouter. You might think U.S. AI companies dictate the AI economy. You’d be wrong. According to data compiled from the OpenRouter model-routing platform, Chinese models consumed over 60% of tokens, while U.S. models fell below 40%. Sure, back in late 2025, the U.S. AI models were riding high. Since then, their decline has been sharp and steep, while their Chinese competitors have charged past them. By June 2026, for example, OpenRouter’s numbers showed Chinese models had reached 18 trillion weekly tokens compared to about 5.5 trillion for American models. Now, these don’t establish as a cold fact that Chinese models process more AI traffic worldwide than U.S. providers. After all, while OpenRouter, a service that lets developers choose among hundreds of models through a common API, serves more than 8 million developers, that’s far from everyone using AI. Still, its numbers are a meaningful read on developer behavior where model switching is relatively easy, and price-performance is exposed. Chinese models first moved ahead of their U.S. rivals in weekly OpenRouter token use in the first two weeks of March, according to an analysis by the China Europe International Business School. Why is this happening? Easy. It’s all about the Benjamins. Or, should I say Chinese Yuans? Chinese developers, including DeepSeek, Alibaba, Moonshot AI, MiniMax, and Zhipu AI, have released increasingly capable models at prices far below those of top closed U.S. offerings, particularly for workloads that generate enormous volumes of output. CEIBS cited DeepSeek V3.2 at $0.42 per million output tokens, compared with $75 per million output tokens for Anthropic’s Claude Opus—a difference that makes Chinese models particularly attractive for high-volume coding, extraction, customer-service, and agent tasks. Additionally, “Chinese models are widely used because they are disproportionately represented in agent workflows run by U.S. companies,” according to OpenRouter COO Chris Clark. They’re winning there because autonomous and semi-autonomous software agents can burn through far more tokens than a consumer chatbot conversation. OpenRouter and Andreessen Horowitz research cited by CEIBS found that programming tasks rose from 11% of model use in early 2024 to more than half, while agent-based automated workflows generated more than half of output tokens. The Chinese companies have also benefited from a strategy that contrasts with the increasingly closed posture of top U.S. model suppliers. Open-weight models can be downloaded, fine-tuned, and served by cloud providers, businesses, or developers themselves. This allows them to spread through repositories, third-party inference services and private deployments. You don’t have to take my word for it, though. According to Clément Delangue, CEO of the open-source repository Hugging Face, Chinese-developed models now account for 41% of all model downloads on his platform. Ironically, all this is happening because of the U.S. government’s export controls on high-end AI chips, which forced Chinese companies to improve their models’ performance on lower-end hardware. As Joseph Hoefer, chief AI officer at Monument Advocacy, said, “It’s possible that restricting Chinese access to top-end chips didn’t just fail to slow capability gains; it may have accelerated the exact efficiency innovation the policy hoped to prevent. If that holds up, the controls are succeeding on the metric they were built to target, compute access, while backfiring on the metric that actually matters, frontier performance.” Good job, Trump! This surge in Chinese model downloads and usage also suggests that for a growing set of production workloads, buyers no longer need the absolute strongest proprietary model. We may still get excited about the hottest new frontier models, but the Chinese models that are good enough, substantially cheaper, and deployable on a customer’s own infrastructure appear to be winning the bulk of everyday AI work. This leaves U.S. frontier providers defending their premium, high end of the market. TECHSTRONG AI PODCAST SHARE THIS STORY