Chinese AI models now handle 57% to 67% of OpenRouter's tokens, up from as little as 6% in February

A token is a unit of volume, and a cheaper model produces more of them for every dollar a customer spends. Token share and dollar share cross at a price that can be solved for. At 57% of OpenRouter tokens, Chinese models take less than half of the spending if their average price per token is below about three-quarters of the other models' ^[1]. At 67%, that line falls to about half ^[1]. Peter Walker, head of insights at OpenRouter, told CNBC the Chinese models are "incredibly cost-effective compared to most models from American labs" ^[10].

The two gateways agree more closely than the ranges suggest. Read at its conservative end, OpenRouter's share rose 44 points, from 13% to 57%, and Vercel's also rose 44 points, from 11% to 55% ^[3]. The generous reading of OpenRouter's range gives 61 points ^[3].

One way this settles is the split Harpreet Arora, head of agentic infrastructure at Vercel, described. "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," he said ^[13]. He added that companies still want frontier US models for some more complicated tasks ^[14]. On that version, US labs lose tokens and keep the higher-priced work.

Another is that the new US models pull volume back ^[11]. Token counts would understate that recovery. Dianne Penn, head of product management, research and labs at Anthropic, told CNBC the company was trying to make its models' answers "more efficient, so it uses less tokens depending on your effort setting" ^[12]. A lab that finishes the same job in fewer tokens shrinks on a token league table.

A third is that capability, more than price, moved the share. Walker said this year's Chinese open-source models "can credibly perform in advanced agentic use cases, especially in regards to coding, in a way that was just not true in late 2025" ^[9].

The export-control strategy in the reporting works on inputs. The US has restricted Chinese AI companies from buying the most advanced chips ^[7], and Washington's further worries are remote access to Nvidia chips through overseas data centers and "distillation," in which new models mimic established ones ^[8]. The usage data measure a later choice: the model a developer picks for a job. The reporting does not show what adoption timeline, if any, the controls were designed around, so it cannot say the switch outran one. AI was a major focus when Donald Trump and Xi Jinping met this week ^[17].

Geography narrows the question. Since the Global South's 67% sits at the top of the platform-wide range, the rest of OpenRouter's traffic, including the US companies that use about half its tokens ^[5], runs at or below the platform average ^[2]. Daniel Remler, a senior fellow at the Center for a New American Security, told CNBC: "The ultimate concern is that the integration of Chinese AI models pulls countries into a Chinese technology sphere of influence that hardens into geopolitical alignment" ^[15].

I think the switch is a pricing decision, made task by task, on models that had already shipped. The counter-thesis is that those models got good through the gaps Washington already names, remote chip access and distillation ^[8]. In that case the chip controls are the relevant lever, and adoption follows from what they failed to stop. The pricing view is wrong if the cheaper US models close much of the price gap and win back little volume.