Read the numbers first, then read how they were produced
Based on OpenRouter data, in the week of August 31 to September 6, total global token usage across large models reached 115 quadrillion, up 1.77% week over week. Chinese models accounted for 56.72 quadrillion, up 2.83%. US models came in at 16.54 quadrillion, down 3.1%. Chinese models have now exceeded US models in weekly usage for nineteen consecutive weeks.
Four of the top five are Chinese. Tencent's Hunyuan Hy4 preview took first place at 14.7 quadrillion tokens, up 379% week over week. OpenAI's GPT-5.6 Luna was second at 12.9 quadrillion, up 66%. Zhipu's GLM-5.3 Flash rose to third at 12.4 quadrillion, up 101%. DeepSeek-V4-Flash in its official release was fourth at 12.4 quadrillion, and the preview build fifth at 5.19 quadrillion. MiniMax M3 returned to the list after nearly a month, sixth at 5.02 quadrillion, up 95%.
Some fell off. Xiaomi's MiMo-V2.5, third the previous week, and Gemini 3.7 Flash, ninth, are both gone.
First, the cold water: this is not global market share
OpenRouter is a third-party model routing platform. What it counts is text request volume inside that platform. There are three built-in biases, and getting them wrong means misreading the whole thing.
Sample bias is the first. The people who route models through OpenRouter are mostly overseas developers, independents and small teams. Large enterprises calling official APIs directly are invisible here. So this ranking measures how popular open-weight and cheap models are among the global long tail of developers. It does not measure who earns the most in the global AI market.
Free tiers inflate the numbers. The sixth entry, MiniMax M3, is literally labeled (free). From August 24 to September 6, developers could use it at no cost through GMI Cloud and OpenRouter. Free things get used more. That is common sense, not proof of capability.
Token volume is not commercial value. A million tokens spent on batch data cleaning and a million tokens spent reviewing high-value contracts carry identical weight in this ranking and differ by an order of magnitude on a P&L.
There is also a separate table by vendor: DeepSeek 23.8%, OpenAI 20.0%, Google 19.8%, Zhipu 9.7%, Alibaba 5.6%, Tencent 3.8%, Anthropic 2.9%. Note that Tencent is only 3.8% there, which does not contradict topping the single-model chart. One Hunyuan model is growing fast, but Tencent's overall footprint is still small. Before reading any ranking, ask whether it counts models or vendors.
Cold water done. Here is the real signal.
Strip out the noise and the list still says three things.
First, the open-weight plus low-price play is genuinely being validated by developers worldwide. Hy4 preview was released and open-sourced on August 28, with 770 billion total parameters, 49 billion activated, and context beyond 1M, tuned for agent, coding and productivity work. A lightweight version followed on September 1, compressing weights from 1.5TB to roughly 214GB, with the team claiming long-document comprehension nearly matches the original. Going from launch to first place took one week, and it happened because three things stacked up: you can self-host it, it is cheap, and it is good enough.
Second, the role of Chinese models is shifting. HUMAIN, the AI company under Saudi Arabia's Public Investment Fund, built its first Arabic large model on top of MiniMax M3. Previously, Chinese models went abroad mainly as APIs or consumer apps. This time one served as the foundation for another country's domestic model. If that path holds, the business changes character: from selling calls to selling the base layer.
Third, turnover on this list is brutal. GLM-5.3 Flash gained 101% in a week. MiMo-V2.5 went from third to nowhere in a week. No position here is safe, and last week's winner can be invisible this week.
What this means for you
If you are choosing a model, how should you use "high call volume" as a signal?
Where it helps: volume usually means a mature ecosystem. Someone has already hit the bugs you are about to hit. Third-party tools integrate with it first. Searchable community solutions exist. For a small team without deep engineering capacity, that can matter more than two or three extra benchmark points.
Where it does not: do not treat it as a performance leaderboard. The list mixes free models, promotional models and long-tail workloads, so movement reflects pricing and distribution strategy more than capability. If you need to pick a model, run your own actual task through it. That beats every ranking.
One more practical point. Since the list is this unstable, do not architect your product around any single model. Abstract the call layer so you can swap. Being able to change models matters more than choosing the right one.