Chinese AI models turn low-cost production into popularity as US models lose ground
- Chinese open-weight AI models have overtaken U.S. models on OpenRouter and Hugging Face usage rankings.
- Chinese models are running more than three times the weekly token volume of American systems after passing them in February 2026.
- Developers and enterprises facing surging bills from OpenAI and Anthropic now have models like DeepSeek, GLM 5.2, and Kimi K3 that deliver comparable results at a fraction of the price.
OpenRouter, the aggregation platform that routes developers’ API calls across dozens of AI systems, reports that open-weight models built by Chinese labs now run more than three times the weekly token volume of American models.
The reason is because Chinese models that are nearly equivalent to their U.S. counterparts cost a fraction as much. Price has become a deciding factor for developers optimizing with AI.
Why are Chinese AI models becoming more popular?
OpenAI and Anthropic sell access to closed models at a premium per token, while labs such as DeepSeek, Moonshot, and Zhipu publish weights developers can download and run without paying a per-token toll.
As of late June, the six most-used models on OpenRouter were all open releases from Chinese firms, including Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai.
For context, in a Cryptopolitan comparison, Zhipu GLM 5.2 runs at $1.40 per million input tokens and $4.40 per million output tokens compared to $5 and $25 for Anthropic’s Opus 4.8.
During the week of February 9 to 15, 2026, Chinese models handled 4.12 trillion tokens on OpenRouter while U.S. models handled 2.94 trillion. This was the first week that the American side of the industry lost. Through most of 2025, U.S. systems had held close to 70% of top-model usage.
Chinese open-weight models are also dominating downloads. Data shows that they made up 41% of Hugging Face downloads this spring, bypassing U.S. releases.
Vercel’s Production Index, published in July, found that open-weight models accounted for 29% of all tokens on its AI gateway in June 2026, up from 11% in April. Those models handled close to a third of tokens while drawing under 4% of spending. DeepSeek alone drove 22.6% of the gateway’s token volume, placing it third behind Google.
Why are industries adopting Chinese technology?
Cryptopolitan reported that a KPMG survey of 2,145 senior leaders found 29% of them cannot understand or control what their AI systems cost to run. Uber burned through its 2026 AI coding budget by April and now caps engineers at $1,500 per tool each month, while one unnamed firm ran up a reported $500 million Claude bill in a single month.
Z.ai’s cloud API, for instance, falls under China’s National Intelligence Law, meaning any code or data sent through it could be accessed by Chinese authorities. U.S. lawmakers opened an inquiry in May into risks from Chinese-origin models in critical infrastructure.
Notably, Moonshot AI is preparing the launch of Kimi K3, a 2.8-trillion-parameter open-source release that the company says competes with top U.S. systems.
On overall benchmarks, the model sits below Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, but Cryptopolitan reported that it beat Claude Opus 4.8 and GPT-5.5 on coding and agent tests. Kimi K3 is priced at $3 and $15 per million tokens, undercutting GPT-5.6 Sol at $5 and $30 and Fable 5 at roughly $10 and $50.
The smartest crypto minds already read our newsletter. Want in? Join them.
FAQs
When did Chinese AI models pass US models on OpenRouter?
The crossover came the week of February 9 to 15, 2026, when Chinese models processed 4.12 trillion tokens against 2.94 trillion for US models, the first time they led on the platform.
How much cheaper are Chinese open-weight models?
Zhipu's GLM 5.2 costs $1.40 per million input tokens and $4.40 per million output, versus $5 and $25 for Anthropic's Opus 4.8, while Moonshot's Kimi K3 at $3 and $15 undercuts GPT-5.6 Sol and Claude Fable 5, per Cryptopolitan.
Do Chinese models actually match US performance?
On several coding tests they do: GLM 5.2 scored 62.1 on SWE-bench Pro against GPT-5.5's 58.6, and Kimi K3 beat Claude Opus 4.8 and GPT-5.5 on coding and agent tasks, though it still trails Claude Fable 5 and GPT-5.6 Sol on overall benchmarks.
Hannah Collymore
Hannah is a writer and editor with nearly a decade of blog writing and event reporting experience in the crypto space. At Cryptopolitan, Hannah contributes to the news page, reporting and analyzing the latest developments in DeFi, RWA, crypto regulation, AI and frontier tech industries. She graduated from Arcadia university with a degree in Business Administration.
