Chinese Models Took 4 of the Top 5 Spots for 27 Straight Weeks—Half Your AI Readers Aren't Using ChatGPT
OpenRouter's weekly token rankings have become the most direct window into the global AI landscape. The data for October 26 through November 1 contains a few numbers worth pulling out on their own.
Total global LLM usage hit 194 trillion tokens for the week, up 3.19%. Chinese models accounted for 99.55 trillion. US models: 20.68 trillion. That's nearly a 5x gap—and the 27th consecutive week Chinese models have come out ahead.
99.55T vs 20.68T
Weekly token volume: Chinese models now process almost five times what US models do. When this first flipped 27 weeks ago, many dismissed it as a blip.
Three signals in the rankings
Chinese models hold four of the top five positions. Here's the breakdown:
| Rank | Model | Weekly tokens | WoW change |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 | 48.1T | +43% |
| 2 | Zhipu GLM-5.3 Flash | 15.0T | +13% |
| 3 | DeepSeek V4.1 Flash (Free) | 14.9T | -17% |
| 4 | GPT-6 Luna | 9.62T | +15% |
| 5 | Tencent HY4 Preview | 9.07T | +1% |
First signal: DeepSeek's 1017 variant is still growing—48.1 trillion tokens, up 43% week over week—while its free tier dropped 17%. Users aren't leaving; they're upgrading to the newer, more capable version.
Second signal: GPT-6 Luna is the only non-Chinese model in the top five, sitting at number four with 9.62 trillion. Growing 15% isn't slow, but in absolute terms it's processing a fifth of DeepSeek 1017's volume. The gap isn't closing. It's calcifying.
Third signal, and the most dramatic: MiniMax M3.1 Flash surged 690% in a single week, jumping from outside the rankings to number eight with 6.23 trillion tokens. Released October 19, it targets agentic tasks, coding, and computer use—nearly 3x faster inference with 70% lower cost. Last week's number eight (Xiaomi MiMo-V2.6-Flash) and number nine (NVIDIA Nemotron3 Ultra) both fell off the chart entirely.
Why Chinese models are winning
Zhu Yibo, CTO of StepFun, put it plainly in a media interview: "The core logic isn't price competition—it's model efficiency and scenario fit winning out."
The usage structure tells the story. The heaviest workloads on OpenRouter come from AI agents like OpenClaw—long-horizon, token-hungry tasks. One agent session refactoring code or running research can burn through tens of thousands of tokens. At that consumption level, unit inference cost determines whether an agent is viable at all.
Chinese models have pushed hardest in exactly this direction. DeepSeek's architecture innovations, MiniMax cutting costs by 70%, Zhipu's lightweight Flash variants—all solving the same problem: making agents affordable to run. China's National Data Bureau reports daily token consumption above 140 trillion, a thousandfold increase from early 2024. JPMorgan projects Chinese AI inference token consumption will reach 3,900 quadrillion by 2030.
What this means for GEO
Most people doing GEO equate "AI search" with ChatGPT or Perplexity and build their optimization around those platforms. OpenRouter's data tells a different story: more than half the world's token consumption happens on Chinese models. When users ask questions through DeepSeek, Doubao, or Kimi, the models reading and citing your content are those—not GPT.
Model differences flow downstream into citation behavior. Training corpora, retrieval preferences, and Chinese-language comprehension depth vary across models. Your content might perform well in ChatGPT but never get cited in DeepSeek—not because the content is weak, but because you never tested it there.
Something concrete to do: take your ten most important business queries and run them through DeepSeek, Doubao, Kimi, and ChatGPT separately. Check whether your brand appears in the citations. Four engines, four answer sheets. If any engine has never heard of you, that's a specific, fixable gap.
Twenty-seven weeks. This isn't news anymore—it's the status quo. The question is whether your GEO strategy has caught up to it.