Chinese AI models rose from 6–13% to 57–67% of OpenRouter tokens

01The line: from a tenth to half or more

CNBC reported on 26 September that Chinese models accounted for 57–67% of OpenRouter tokens in the week of 14 September, up from 6–13% in February. The range covers three groups of OpenRouter’s business customers: the US, Europe, and what OpenRouter calls the “Global South”, 82 countries in Central and South America, Africa and Asia. The chart CNBC published with the article shows that the February figure is February 2025, and it ends at the week of 7 September 2026, when the share was 68% in the Global South, 50% in the US and 48% in Europe. About half of OpenRouter’s tokens are used by US companies.¹ On Vercel’s gateway, the Chinese share rose from 11% in January to 55% in August.¹

OpenRouter’s own public weekly series, which we pulled, show the same climb, and the total under it grew from about 5.7 trillion tokens a week at the start of November 2025 to about 132 trillion in the week of 21 September.³⁴

One number does not line up. In July CNBC put the US weekly peak since 8 February 2026 at 46%; its September series puts several June weeks for the US above 55%.²¹ Neither article explains why, so we quote CNBC’s September figures as reported.

02What drives it: price first, then licence and quality

Price. “Chinese models are becoming capable enough for more tasks at a much lower cost. Once a model meets the quality bar for the job, that price difference becomes compelling,” Harpreet Arora, head of agentic infrastructure at Vercel, told CNBC; he added that companies still want US frontier models for more complicated tasks.¹ In July, Justin Summerville of OpenRouter put open Chinese models at “60% to 90% cheaper” than the leading Anthropic and OpenAI models.² The spend data agree: in the week of 7 September, Chinese models took 48–68% of tokens by region but 26–33% of what customers spent.¹

Licence and control. “We’re seeing companies increasingly motivated to turn to cheaper AI stacks they can control and adapt themselves,” Yacine Jernite, head of machine learning at Hugging Face, told CNBC in July.² DeepSeek published V4.1 Flash on 10 September under the MIT licence.⁶ Z.ai’s GLM-5.3 licence allows commercial use but requires a firm running a model-as-a-service business with more than US$10bn of revenue in 12 months to pass Z.ai’s security review first.⁷ MiniMax and Moonshot AI publish their current models under licences of their own.⁸

Quality. Peter Walker, head of insights at OpenRouter, said Chinese open models released this year “can credibly perform in advanced agentic use cases, especially in regards to coding, in a way that was just not true in late 2025.”¹ Kyle Chan of Brookings put them “six to nine months” behind in July.² The other side of the argument is in the same reports: Daniel Remler of the Center for a New American Security called Chinese AI “real economic and security risks for the United States”, and two US House committees are investigating its adoption.¹

03What it means when a team picks a model

• Cost. The spend gap is the case for testing a cheaper model per task; the same reports say teams still send harder work to frontier models.¹ Price per token is not cost per task: a model that needs more tokens or more retries to finish can cost more.
• Where the data goes. Where a model was built and where a request is processed are separate questions. On a gateway, one open-weight model can be served by several hosting companies; OpenRouter lets a customer exclude named hosts, route only to hosts that do not store prompts, or require zero data retention.⁹ Those settings are off unless someone turns them on.
• Governance. Keep the model’s name and version, its licence, the host and the host’s retention terms on file; each can change on its own, and so can the rules, as Washington’s inquiry shows.¹