Chinese Models Take 30%+ of US OpenRouter Token Use Since Feb 8

A single data point from OpenRouter is doing more to describe where US AI spending is actually going than most of this quarter's earnings commentary. CNBC reports that the share of tokens US companies run through Chinese-origin AI models on the platform has sat above 30% every week since February 8, and has climbed as high as 46%. The average across the prior twelve months was 11%. In the first half of 2025 it was only 4.5%.

The reason is cost, and the gap is not incremental. Justin Summerville, who works on data and analytics at OpenRouter, told CNBC that open source Chinese models can run 60% to 90% cheaper than the leading Anthropic and OpenAI models. The per-workload figures in the piece are blunt: Anthropic's Claude at $4,811, OpenAI's ChatGPT at $3,357, and Zhipu's GLM at $544 for similar work. Vercel's tracking of Z.ai's GLM 5.2, released in June, showed the fastest adoption of any model on its platform in 2026, with daily token volume growing about 27x and customer count about 80x in the model's first full week after launch.

Why this matters if you are pricing, benchmarking, or budgeting AI products: the story the frontier labs tell investors is that scale plus quality justifies premium tokens. What the OpenRouter data suggests is that a real chunk of US token demand is now behaving like a commodity, routed to whichever open model is good enough for the task. Summerville's framing in the CNBC piece is that when a task does not need the best model, teams are beginning to route it to the cheapest one that is good enough, and the recent wave of Chinese models is winning that trade.

The honest caveat is that OpenRouter is one venue, not the whole market, and the reporting does not break out how much of that 30 to 46% is production enterprise traffic versus developer experimentation, or which workload categories are actually moving. What is not in dispute is that a category the US labs assumed they owned is being priced away from them. The vendors positioned to gain are the routing and observability layers, and any US buyer whose cost model was written when cheap capable open weights did not yet exist.