Chinese Models Just Peaked at 46% of US Token Traffic. My Routing Math Isn't Surprised.

CNBC put a number on July 7 that would have sounded like a typo a year ago: since February 8, the share of tokens US companies route to Chinese models through OpenRouter has held above 30% every single week — peaking at 46%. The average over the twelve months before that was 11%. In June I wrote an open-weight showdown arguing GLM-5.2 and MiniMax M3 were the best budget models most builders hadn't priced in. This is the market grading that homework at scale.

The numbers

The pull is coming from three names: Zhipu's GLM-5.2, DeepSeek, and Moonshot's Kimi. GLM-5.2 is the sharpest data point — on Vercel's infrastructure alone, its daily token volume grew 27-fold in its first week of availability, and it landed within a percentage point of Claude Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost. Across the category, CNBC quotes the discount at 60 to 90% below the leading Anthropic and OpenAI models.

Price is doing the work

The quote in the CNBC piece that explains all of it: "price is doing the work here." Teams aren't converting to an ideology; they're routing to the cheapest model that's good enough, and the current wave out of China is winning that trade for a growing set of workloads. And it isn't just startups arbitraging OpenRouter — Microsoft has started swapping its own MAI models in for OpenAI and Anthropic frontier models inside Excel and Outlook production workloads. Single-model loyalty is dead at every scale; the hyperscalers just call it vendor diversification.

None of this is new logic — it's the same math I did when Gemini 3.5 Flash beat Opus 4.7 on agent benchmarks at a third the price. What's new is the volume: routing-by-price has gone from a budget builder's trick to a third-to-half of a US routing platform's traffic.

The asterisks I'd read before moving traffic

Where the tokens run matters more than where the weights were born. Open weights served by a US or EU inference provider is a very different risk profile from shipping prompts to an API endpoint in Beijing. On OpenRouter, check the actual provider behind the model — the label "Chinese model" conflates the two.

A benchmark point is not your workload. "Within a point of Opus 4.8" on one agentic benchmark tells you the model is worth testing, not worth defaulting. Run your own evals on your own tasks before a single production token moves.

Open weights are only sovereignty if someone can serve them. A frontier-scale MoE is not a one-VPS deployment. For most of us, "open" means "multiple competing hosts" — which is still worth a lot, because nobody can switch off weights the market can re-host. I watched a government switch off my closed-weight default model in June, so I don't treat that property as theoretical.

The whiplash cuts both ways. The same geopolitics that suspended Fable 5 could just as easily regulate Chinese-origin models inside US products. If half your traffic runs through them, that's a dependency with foreign-policy exposure. Keep the fallback aliased — same rule as always.

My actual routing table

What moves to the GLM-5.2 class: bulk summarization, tagging and classification, first-draft prose, low-stakes batch jobs — anywhere volume is high and a miss is cheap. What stays on Claude: agentic coding, anything holding tool access to my server, security review — anywhere a miss is expensive and blast radius is real. What's in testing: long-context retrieval work, where the cost gap is most tempting and the quality variance is widest.

The verdict

The 46% headline isn't ideology; it's invoices. Route by price where quality clears your bar, hold the line where blast radius lives. And read the headline with its own asterisk: 46% is a peak on one router's telemetry, not the whole market — but 30%-plus for five straight months is a trend, and pretending otherwise is just paying more for the same tokens.

Sources

The primary report is CNBC's July 7 piece on the OpenRouter data. MLQ breaks down the cost-gap numbers, and Yahoo Finance has the enterprise-adoption angle, including the GLM-5.2 growth figures on Vercel.