Chinese open-weight models now handle somewhere between 30% and 46% of the enterprise API token traffic flowing through US developer platforms. That is not a projection or a survey sentiment. It is measured gateway data, confirmed by CNBC on July 7 using numbers from OpenRouter and Vercel — two of the platforms that actually route the tokens.
A year ago the same figure was 11%. In the first half of 2025 it was as low as 4.5%. Through OpenRouter, the Chinese share has stayed above 30% of all gateway tokens every single week since February 8, 2026. The shift is not coming. It has already happened, and most Western coverage is still treating it as a hypothetical.
The instinct is to read this as a geopolitics story. It isn’t. It’s a pricing story, and the pricing story is one Western labs wrote themselves.
THE NUMBERS ENGINEERS ARE ACTUALLY LOOKING AT
Start with the model that made the trend concrete. Z.ai’s GLM-5.2 scored 62.1% on SWE-bench Pro — above GPT-5.5’s 58.6% — at $1.40 input and $4.40 output per million tokens. It carries an MIT license whose terms explicitly say “no regional limits.” On one closely watched agentic benchmark, it landed within a single percentage point of Anthropic’s Opus 4.8 at roughly a fifth of the cost.
The adoption curve that followed was not gradual. In GLM-5.2’s first full week on Vercel, daily token volume grew about 27 times and the number of customers using it grew about 80 times. Vercel’s head of agentic infrastructure confirmed those figures directly. Eighty-times customer growth in a week is not a normal adoption pattern for any model on the platform. It is what happens when a price floor collapses and the thing underneath it is good enough.
Justin Summerville at OpenRouter put the gap plainly: open-source Chinese models run 60 to 90 percent cheaper than leading Anthropic and OpenAI models. When the discount is that large and the capability gap is five to ten points on tasks that don’t need frontier performance, the routing decision makes itself.
THE ADVISOR MODEL PATTERN
What’s actually driving the volume isn’t wholesale replacement. It’s a routing pattern that went mainstream in the second quarter: a cheap open-weight model handles the bulk of tasks, and escalates to a frontier model only when the task genuinely demands it. Practitioners have started calling it the advisor model. The cheap model does the reading; the expensive model does the deciding, and only when deciding is hard.
Once that pattern is in place, the default tier stops being a Western frontier model and becomes whatever is cheapest that clears the quality bar. Right now that is increasingly a Chinese open-weight model. Routine summarization, code completion, data extraction, first-draft customer support — none of these need Opus-class reasoning, and paying Opus-class prices for them stopped being defensible the moment a 62%-on-SWE-bench alternative showed up at a fifth of the cost.
Yacine Jernite at Hugging Face framed the deeper motivation: companies increasingly want a stack they can control and adapt themselves, and given the current state of open-weight models, that often means Chinese options. The Fable 5 export-control episode in June sharpened that instinct. When a US model can be switched off for everyone by government order on three days’ notice, an MIT-licensed model with “no regional limits” starts to look less like a discount and more like insurance.
THE PRICE TRAP WESTERN LABS BUILT
Here is the part that should concern the frontier labs. They walked into this.
Across 2026, Western labs raised prices at the top. GPT-5.5 doubled GPT-5.4. Fable 5 shipped at $10 input and $50 output per million tokens. Gemini 3.5 Flash tripled its predecessor. Each individual decision was rational — frontier capability is expensive to serve, and the top of the market pays for capability regardless of cost. But the aggregate effect was to vacate the middle. Western labs raised prices into the exact window when Chinese open-weight models reached near-frontier performance at deeply discounted prices.
The result is a market splitting cleanly in two. Western labs are defending the top tier, where a task is hard enough that a six-point performance gap is decisive and cost is secondary. That tier is real and probably defensible. But the middle tier — the vast volume of enterprise workflows where “good enough and cheap” beats “excellent and expensive” — is being ceded. And the middle tier is where most of the tokens live.
There is a genuine counter-consideration, and it is not small. Direct API calls to Chinese providers route through Chinese-jurisdiction servers, which typically violates data-residency requirements for regulated industries — healthcare, finance, legal. There are content restrictions on politically sensitive topics and reported gaps in tool-call schema reliability. For a bank or a hospital, the jurisdiction problem alone can disqualify the cheaper option regardless of price. The routing to Azure-hosted or Cloudflare-hosted versions of these models exists precisely to thread that needle, and how well it works will determine how much of the regulated middle tier is actually reachable.
WHAT THIS MEANS IF YOU BUILD
The takeaway for anyone architecting an AI product is not “switch to Chinese models.” It is that the era of defaulting every call to a single frontier vendor is over, and the economics have made that decision for you whether you’ve formalized it or not.
The teams handling this well have already separated their traffic by tier: a cheap default model for the bulk, a frontier model for the hard exceptions, and a routing layer that decides between them per task rather than per contract. They’ve made the model name a swappable config value, not a load-bearing assumption. That same discipline is what made the Fable 5 recall a non-event for the teams that had it, and it is what makes the Chinese-model price advantage capturable without betting the whole stack on a single jurisdiction.
The 30-to-46% figure is not the end state. It is a midpoint in a correction that is still running. Western frontier labs will keep the top of the market. The open question — the one worth watching through the back half of 2026 — is how much of the middle they can win back, and whether they even try, or whether they decide the top tier alone is the business they want to be in.