In-Depth Analysis of Financial News: The Second Half of the AI Model “Invasion” – Chinese Models Making a Comeback with “Value for Money”
Summary of Key Points
Over the past two weeks, the global AI model industry has entered a period of rapid iteration, with updates occurring almost weekly. Although overseas giants such as Anthropic, OpenAI, and Meta still top the list in terms of pure technical capabilities, the market landscape has undergone a subtle yet significant shift: Chinese models have surpassed those from the United States in terms of actual usage by developers for 20 consecutive weeks, accounting for more than 70% of the traffic on the OpenRouter platform.
The core reason for this phenomenon is a shift in the industry's focus from simply pursuing the “most powerful” models to emphasizing “Token ROI” (the cost-effectiveness of using AI models). For the vast majority of daily development and application scenarios, Chinese models provide sufficient performance at a fraction of the cost of their overseas counterparts. This combination of “sufficiency and affordability” has made Chinese models the preferred choice for businesses and developers, especially as the cost sensitivity increases with the widespread adoption of AI in business processes.
---
Detailed Explanation
1. Rapid Pace of Innovation: Model Updates from Annual to Weekly
Previously, model upgrades were a slow process, possibly occurring only once every six months to a year. However, in the past two weeks (September 1–11), seven leading companies from both China and abroad released new models, including Anthropic’s Claude Fable 5.1, OpenAI’s GPT-6 Astra, Google’s Gemini 3.8 Flash, as well as domestic models like Qianwen, DeepSeek, and Kimi.
What does this mean?
It indicates that the AI competition has intensified, with model iteration outpacing the ability of developers to evaluate and switch between different models. In the past, it might have taken half a year to switch to a new model; now, a cheaper and faster model could emerge just the next month. This rapid iteration brings significant uncertainty to the industry, forcing users to remain highly vigilant and ready to switch tools at any time to obtain the best solution.
2. A Contradictory Ranking: Overseas Models Lead in Capability, but Chinese Models Dominate in Usage
There is a striking contrast in the rankings:
- By Capability (Intelligence Test): The top five models on the list of model intelligence are all overseas. Anthropic’s Claude Fable 5.1 ranks first, followed by OpenAI’s GPT-6 Astra. China’s best model, GLM-5.3, ranks seventh, and Kimi K3 ranks ninth. This shows that overseas models still hold a technological advantage in solving complex and advanced tasks such as logical reasoning and code generation.
- By Usage (Actual Traffic): On the OpenRouter API aggregation platform, Chinese models have led in usage for 20 consecutive weeks, accounting for more than 70% of the total traffic. A year ago, only two Chinese models were in the top ten; now, they have almost completely dominated the list.
Why is this the case?
Most people do not need the “most powerful” models; they just need models that are “sufficient” for their needs. It’s like buying a car: while a Ferrari is fast, 99% of people would prefer a fuel-efficient and durable Toyota or比亚迪 for daily commuting. Chinese models have precisely targeted this market segment.
3. The New Focus: From “Competing in Intelligence” to “Competing in Value for Money”
Analysts at UBS Securities note that in the second half of 2026, the industry’s focus has shifted from “Token-maxxing” (using as much AI as possible) to “Token ROI.”
In simpler terms:
- First Half of the Year: Many companies encouraged unlimited use of AI, believing that more usage meant greater advancement, but this led to high costs and difficulty in measuring the actual benefits.
- Second Half of the Year: Managers began to reconsider. They realized that expensive overseas models were overkill for tasks like writing code, processing documents, and simple data analysis. Thus, the focus shifted to “Token Optimization” – using models that provide the best quality at the lowest cost per use.
Advantages of Chinese Models:
The API prices of domestic models are typically 10% to 20% of their international counterparts. For example, DeepSeek V4.1 Flash is very cost-effective, with some tasks costing only 1/37 to 1/76 of the cost of top-tier models. This cost difference can be crucial for startups or companies deploying AI on a large scale; using overseas models could deplete their capital, while Chinese models enable scalable implementation.
4. Real-World Examples: Developers Making Decisions Based on Practical Use
The story of the open-source project xiaobei’s developer is illustrative:
- At the beginning of the year: He used the expensive Claude as his primary model, only switching to domestic models for simpler tasks.
- Since May: He has switched to all domestic models (GLM 5.3, Alibaba Qianwen, DeepSeek V4.1 Flash).
- Reason: He explained, “99.99% of programming tasks are not challenging enough for models like GLM 5.3.” Unless dealing with extremely complex tasks like developing operating system kernels or rendering engines, there’s no need to pay a much higher price for top-tier models.
Market Trends:
More leading global IT companies are incorporating Chinese open-source models into their workflows. This is not because Chinese models have surpassed overseas models in every aspect, but because they have achieved a perfect balance in price, speed, and stability for most use cases.
5. The True Advantage of Low Prices: Efficiency, Not Losses
Some might wonder if Chinese models are sold at a loss. UBS’s research indicates that Chinese model manufacturers have a gross margin of around 20% to 40% on their API services, indicating they are not sacrificing profits to expand the market.
What does this mean?
Chinese manufacturers have reduced costs through innovations in model architecture and optimizations in training and inference efficiency. They have gained competitiveness through technical excellence, not through subsidies. DeepSeek V4.1 Flash can process 1 trillion tokens in 24 hours and is expected to handle 2.8 trillion tokens in 48 hours, demonstrating its high throughput and low cost per token.
Conclusion:
Chinese models have not won the “technological peak” battle, but they have secured a significant share of the “large-scale application” market. With their excellent value for money, they have become the default choice for AI implementation. As AI moves from being a novelty to a daily necessity, the market share held by cost-sensitive users will only continue to grow. Overseas models will continue to dominate niche markets for high-end applications, while Chinese models will dominate the broader, more general market.