DeepSeek Leads: Chinese AI Models Record 21st Consecutive Week as Global #1 in Token Usage

Overview

According to the latest OpenRouter data, Chinese AI models processed 67.46 trillion (T) tokens last week, while US models processed only 14.21 trillion tokens. This vast gap marks the twenty-first consecutive week that Chinese models have maintained the global lead in token usage.

More notably, four of the global top five models are held by Chinese manufacturers, demonstrating the Chinese artificial intelligence industry's significant advantages in application scale and user stickiness. DeepSeek, bolstered by the strong performance of its V4.1 Flash version, has become the absolute mainstay for token consumption this week.

Leaderboard Landscape: Chinese Models Sweep the Global Top Five

In the latest global token usage rankings, Chinese models display an overwhelming cluster advantage:

• DeepSeek V4.1 Flash leads with 15.8 trillion tokens, a 219% increase over the previous period
• GLM5.3 Flash ranks second, processing 14.1 trillion tokens
• Tencent Hunyuan Hy4 Preview ranks third, reaching 12.5 trillion tokens
• #5 DeepSeek V4 Flash 0731 also enters the top five, consuming 9.44 trillion tokens

This leaderboard structure stands in sharp contrast to the historically US and Europe-dominated landscape. The DeepSeek series occupies two spots, with total consumption far ahead, reflecting its extensive penetration within developer communities and enterprise-level applications.

"Token usage is the most direct indicator for measuring the actual scale of AI model adoption. Chinese models leading for twenty-one consecutive weeks indicates that they have achieved breakthroughs not only technically but also won users in real-world scenarios."

Technical Breakthroughs and the Cost-Effectiveness Paradox of DeepSeek V4.1 Flash

DeepSeek V4.1 Flash was released on September 10, adopting a 552 billion parameter Mixture of Experts (MoE) architecture with native image understanding capabilities. It is optimized for KV Cache size tailored to Agent workloads, aiming to reduce inference costs.

However, a lower unit price does not equate to lower task costs. According to practical data from Artificial Analysis:

1. V4.1 Flash averages approximately 89K tokens output per task
2. In contrast, V4 Flash 0731 averages only 62K tokens output per task
3. This means V4.1 Flash consumes significantly more tokens per single task

Despite having a more competitive price per token, due to the increased complexity of individual tasks, the actual task cost of V4.1 Flash is slightly higher than that of V4 Flash 0731. This phenomenon reveals a key trend in large model applications: users are no longer choosing merely the "smartest model," but rather the "most cost-effective model capable of delivering stable results."

Industry Insights: From Technical Competition to Scalable Deployment

The sustained leadership of Chinese AI models in token usage reflects an important shift—the focus of the artificial intelligence competition is moving from "who possesses the most powerful model" to "who can enable users to adopt the model on a large scale and at high frequency."

DeepSeek's success stems not only from technical architecture innovation but also from its precise positioning of cost-sensitive user groups, such as independent developers and small to medium-sized enterprises. By democratizing high-end capabilities through Flash versions and performing targeted optimizations for Agent scenarios, DeepSeek has successfully constructed a complete user closed loop ranging from research to production.

In the future, as MoE architecture and inference optimization continue to evolve, token usage is expected to remain a critical indicator for measuring the commercial value of AI models. Whether Chinese models can sustain this momentum is worth continued attention.