Chinese AI models now handle more than 30% of OpenRouter's weekly token traffic, up from 11% a year ago

Since February 8, Chinese-developed AI models have taken more than 30% of the tokens routed through OpenRouter every single week, peaking at 46%. A year ago the weekly average was 11%. That is a fast move, and it is worth being precise about what it does and does not show.

What the number actually counts

OpenRouter is a routing marketplace. Developers point their app at it, and it forwards each request to whichever model they pick from dozens of providers, with no lock-in to any one lab. Because the traffic is real production usage rather than a benchmark score or a press-release claim, the token totals are a decent read on what developers reach for by default. They are also a slice, not the whole market. OpenRouter is one router used by a particular crowd of developers and startups. It is not enterprise-wide AI spend, and a company that calls OpenAI or Anthropic directly, or runs models on its own hardware, never shows up in it.

Two different figures are circulating, and they measure different things. The first, from CNBC citing OpenRouter, is that tokens sent to Chinese models by US companies have stayed above 30% weekly since February. The second, from a Bloomberg chart built on OpenRouter and Exponential View data, is that the share of tokens going to US-built models (Google, OpenAI, Anthropic combined) fell from about 70% in June 2025 to about 30% in June 2026. One tracks Chinese usage climbing; the other tracks the American share falling. They point the same direction, but they are not the same stat, and neither is a measure of total US AI consumption.

Price is doing the work

The reason is cost, and nobody involved pretends otherwise. Justin Summerville of OpenRouter told CNBC that Chinese open-weight models run 60% to 90% cheaper than the leading US options. Tencent's Hy3 Preview, for one, lists at $0.063 per million input tokens, which is the kind of number that changes an invoice when you are serving traffic at scale.

The switching logic is not complicated. When a task does not need the strongest available model, teams route it to the cheapest one that clears the bar. The startup Lindy moved all of its traffic off Anthropic's Claude to DeepSeek; CEO Flo Crivello said the change saves the company millions. Z.ai's GLM 5.2, out in June, saw the fastest adoption of any model Vercel tracked this year, with daily token volume up roughly 27x and its customer count up about 80x in the first full week after launch. Adoption that sticks after the launch spike is the tell that these models are getting wired into pipelines, not just tested and dropped.

On OpenRouter's June leaderboard, DeepSeek V4 Flash led all models by volume, Tencent's Hy3 Preview sat second, and Chinese open-source models held six of the top ten spots. DeepSeek alone accounted for 16.3% of all tokens on the platform, ahead of any single US provider.

The part the token chart does not settle

Cheaper and more-used is not the same as better, and the capability picture is less lopsided than the traffic. In May, the US government's Center for AI Standards and Innovation put the leading Chinese models about eight months behind the top US models across cybersecurity, software development, math, science, and abstract reasoning. Kyle Chan at the Brookings Institution lands in the same range, six to nine months. Anthropic's Dario Amodei has argued that Chinese models are tuned for benchmarks and distilled from US labs, and that raw capability, not price, decides adoption over the long run.

The data gives Amodei one datapoint to stand on. Claude Opus 4.7 holds third place on the OpenRouter leaderboard, the highest-ranked closed model, and Claude's usage has held up across several rounds of Chinese releases. That durability is the exception, though. For the rest of the US field, the token share is going one way.

Read plainly, the OpenRouter numbers describe a routing decision, not a verdict on which models are strongest. Developers are sending the work that does not need a frontier model to whatever runs it for less, and right now that is increasingly a Chinese open-weight model. That is a real shift in defaults, worth watching, and it is a narrower claim than "China is winning AI." The token chart measures where the cheap-enough work goes. The capability benchmarks, which still favor the US labs by most of a year, measure something else.