Something interesting is happening in the AI model market.
Looking at this week's OpenRouter model usage ranking, 8 of the Top 10 models by token usage are Chinese models.
Yes — 8 out of 10.
Here is the current ranking:
| Rank | Model | Weekly Tokens |
|---|---|---|
| 1 | Tencent: Hy4 Preview | 20.4T |
| 2 | OpenAI: GPT-5.6 Luna | 14.8T |
| 3 | DeepSeek: DeepSeek V4 Flash 0731 | 12.9T |
| 4 | Z.ai: GLM-5.3 Flash | 12.7T |
| 5 | DeepSeek: DeepSeek V4 Flash 0423 | 5.15T |
| 6 | Tencent: Hy3 | 4.01T |
| 7 | NVIDIA: Nemotron 3 Ultra (free) | 3.82T |
| 8 | Xiaomi: MiMo-V2.5 | 3.82T |
| 9 | Z.ai: GLM-5.3 | 3.48T |
| 10 | Z.ai: GLM-5.2 | 2.51T |
The only two non-Chinese models in the Top 10 are:
Everything else comes from:
And there is another number that makes this even more interesting.
If we add up the weekly usage shown in the ranking, the Top 10 models account for approximately:
83.59 trillion tokens
The eight Chinese models account for approximately:
64.97 trillion tokens
That's roughly:
78% of all token usage represented by the Top 10.
So Chinese models aren't just occupying many positions on the leaderboard.
They're also representing the majority of actual token volume within this Top 10.
That's a much more interesting signal.
Perhaps the biggest surprise is the model at the top.
Tencent Hunyuan Hy4 Preview currently sits at #1 with:
20.4 trillion tokens
That's significantly ahead of OpenAI GPT-5.6 Luna at 14.8T.
Hy4 Preview is Tencent's latest large MoE model, with:
What's interesting is that Hy4 isn't positioned as a simple chatbot model.
It's designed around a broader shift:
Question
↓
Answer
is becoming:
Goal
↓
Plan
↓
Use tools
↓
Execute
↓
Observe
↓
Adjust
↓
Continue
↓
Complete task
That kind of workload can also consume a lot of tokens, which is worth remembering when interpreting usage rankings.
DeepSeek occupies two positions in the Top 5:
12.9T tokens
5.15T tokens
Combined, that's more than:
18 trillion tokens in one week
across these two V4 Flash versions alone.
DeepSeek has become one of the most recognizable Chinese AI brands globally, but this ranking shows something more important than awareness:
developers are actually using the models at scale.
And the Flash family illustrates an important trend in AI infrastructure.
Not every workload needs the largest and most expensive frontier model.
For many production applications, developers care about a balance between:
Intelligence
×
Speed
×
Cost
×
Reliability
A model that is slightly weaker on a benchmark but dramatically cheaper to run can be much more attractive at scale.
Z.ai may actually have the most interesting representation in this ranking.
Three GLM models appear in the Top 10:
GLM-5.3 Flash alone processed:
12.7T tokens
That puts it almost level with DeepSeek V4 Flash 0731.
The recent GLM direction is also interesting because the models are becoming increasingly focused on:
GLM-5.3 Flash is particularly notable because it combines:
320B total parameters
18B activated parameters
with a highly efficient hybrid attention architecture.
This is another trend we're seeing across Chinese AI labs:
The race isn't only about making models bigger. It's also about making intelligence cheaper to run.
A year ago, when many international developers talked about Chinese AI, the conversation often started and ended with:
DeepSeek.
That picture is changing quickly.
Look at this Top 10 again:
Tencent
DeepSeek
Z.ai
Xiaomi
Four different Chinese companies are represented.
And other major Chinese AI ecosystems include:
This matters because we're not looking at one breakout model anymore.
We're looking at an increasingly broad AI model ecosystem.
Different companies are competing on different dimensions:
Coding
Reasoning
Agents
Multimodal
Long context
Latency
Cost
Tool use
Productivity
That creates much more choice for developers.
There is an important caveat.
This is a usage ranking, not an intelligence leaderboard.
20T tokens does not mean a model is objectively better than one processing 5T tokens.
Token usage can be influenced by many factors:
For example, an agent that reads an entire repository and performs dozens of tool calls can consume far more tokens than a chatbot answering simple questions.
A cheap model may also be used much more aggressively than an expensive frontier model.
So the conclusion shouldn't be:
"Chinese models are better because they use more tokens."
The more interesting conclusion is:
Chinese models have clearly moved from being alternative models to models developers are actively using at very large scale.
That's an important distinction.
One thing Chinese AI providers have been especially aggressive about is API pricing.
This changes how developers think about model selection.
Suppose one model is 5% better for your workload but costs 10× more.
Which one should you use?
For a low-volume application, perhaps the stronger model.
For billions of tokens of production traffic, the answer may be very different.
This is why I think the most useful model metric is increasingly not:
Benchmark score
or even:
Price per million tokens
but:
Cost per successfully completed task
A coding agent might use a more expensive model but finish in fewer attempts.
A cheaper model might consume more tokens but still have a lower total cost.
The economics depend on the workload.
Another thing that stands out in this ranking is the popularity of Flash-class models.
We have:
near the very top.
That makes sense.
Production AI workloads often need:
Good enough intelligence
+
Low latency
+
Low cost
+
High throughput
rather than maximum benchmark performance on every request.
For example, an application might use:
Simple tasks
↓
Flash model
Medium tasks
↓
General model
Difficult tasks
↓
Frontier reasoning model
That's likely to become a common architecture.
The question isn't necessarily:
Which model should my application use?
It may instead become:
Which model should handle this particular request?
This ranking is also another reminder of how quickly the model market changes.
Today:
Hy4 Preview
DeepSeek V4
GLM-5.3
GPT-5.6
MiMo
are receiving huge amounts of traffic.
Next month, the ranking may look completely different.
That's why I think tightly coupling an application to a single AI provider is becoming increasingly limiting.
A more flexible architecture could look like:
Application
↓
AI Model Layer
↓
├── GPT
├── Claude
├── Gemini
├── DeepSeek
├── Qwen
├── GLM
├── Hunyuan
├── MiniMax
└── Others
Then select models according to:
The more competitive the model market becomes, the more valuable this flexibility becomes.
One of the things we're trying to solve with ApiHub is making this growing model ecosystem easier for developers to access.
Several of the Chinese models appearing in this ranking are already available through ApiHub, including models from:
Instead of creating a completely separate integration every time you want to test another model, ApiHub provides a unified way to access and compare different models.
We support multiple integration styles:
And we provide free credits so developers can test different models before deciding what works best for their application.
The goal isn't to tell developers:
"Model X is the best."
It's to make it easier to answer:
"Which model is best for my workload?"
by actually testing them.
The headline is surprising:
8 of OpenRouter's Top 10 most-used AI models this week are Chinese.
But I think the bigger story is underneath it.
Chinese AI models are increasingly competing on:
And developers are actually using them.
At the same time, OpenAI, Anthropic, Google, NVIDIA, and other labs continue pushing the frontier forward.
That's good for developers.
More competition means:
More models
↓
More choices
↓
Lower prices
↓
Faster iteration
↓
Better AI applications
We're moving away from a world where choosing an AI model meant choosing between only two or three companies.
The model layer is becoming a competitive marketplace.
And that may ultimately matter much more than who happens to be #1 this week.
What do you think?
Are Chinese AI models becoming part of your default model stack, or do you still mainly use GPT, Claude, and Gemini?
And if you've tried Hy4, DeepSeek, GLM, Qwen, or other Chinese models, which one surprised you the most?
Disclosure: I'm building ApiHub, a unified AI API platform designed to make multiple AI models easier to access, test, compare, and integrate.