Chinese AI Models Lead Global OpenRouter Usage as Weekly Traffic Hits 113 Trillion Tokens

A National Business Daily analysis of OpenRouter data for August 24–30, 2026, described a decisive week for China’s AI model ecosystem. Global traffic reached 113 trillion tokens, up 21.11% from the previous week, while Chinese models reportedly led U.S. models in inference volume for an 18th consecutive week. 13

The headline is especially notable because the leading model was not initially marketed under a recognizable company name. The anonymous Ox Alpha—also called “Niu Lai” by some Chinese developers—was later identified by Z.ai, also known as Zhipu AI, as GLM-5.3-Flash. 513

China widened its lead in reported inference volume

The analysis attributed 55.16 trillion tokens to Chinese models, compared with 17.07 trillion for U.S. models. It also described the week as the first time the top three models were all Chinese-developed systems. 13

That distinction matters, but OpenRouter traffic should be read as a measure of usage flowing through a global model-routing platform—not as a direct census of all AI use worldwide. The figures show where traffic on that layer was concentrated during the measured week; they do not, by themselves, explain the users’ locations or every workload behind the tokens.

Ox Alpha became GLM-5.3-Flash—and took the top spot

Ox Alpha reportedly ranked first with 15.7 trillion tokens. DeepSeek-V4-Flash followed with 12.3 trillion, while Xiaomi MiMo-V2.5 remained in third place for a second consecutive week at 9.14 trillion tokens. 13

The model first appeared on OpenRouter on August 20 as an unnamed, free-access listing. Reports described a context window of roughly one million tokens and support for text, image and video input. Z.ai confirmed the model’s identity on August 26, then released open weights under an MIT license. 257

The launch strategy helped turn a model mystery into a large-scale real-world test. Developers could try the system without the normal price barrier, while its long context and multimodal input made it relevant to coding and extended agent workflows. OpenRouter later described Ox Alpha as its biggest model launch, with more than 20 trillion tokens processed in six days. 15

Low pricing was part of the adoption story

After the free preview, reported standard pricing for GLM-5.3-Flash was $0.15 per million input tokens and $0.50 per million output tokens. A launch promotion cut those rates to $0.075 and $0.25 through September 9, according to Z.ai and OpenRouter posts. 115

The precise comparison with GLM-5.3—described in the original reporting as roughly one-tenth the price, or one-twentieth during the promotion—was not independently established by the supplied evidence. The broader point is better supported: inexpensive access, a free trial and high-capacity routing can produce rapid token growth even before a model has built a long public track record.

Usage rankings are not the same as quality rankings

The reporting also cited strong DeepSWE results for GLM-5.3-Flash against several named models. Those numbers should be treated as benchmark-specific claims, not as a definitive ranking of general model quality. The available evidence includes conflicting figures for the benchmark, and a coding score cannot measure every capability or production constraint. 4

For developers, the practical lesson is to separate three questions:

1. How much traffic is a model attracting?
2. How does it perform on the workload that matters?
3. What does each successful task cost after routing, caching and output usage?

OpenRouter’s weekly data speaks most directly to the first question. It suggests that developers were willing to move substantial workloads toward a newly introduced Chinese model when the access conditions were attractive.

Other models show a more competitive market

The same analysis reportedly placed Gemini 3.7 Flash at 3.95 trillion tokens, with growth of 120%, while GLM 5.2 and DeepSeek-V4-Pro-0423 left the leading group. These specific figures were not independently corroborated in the supplied sources, so they should be viewed as reported rather than settled platform-wide measurements.

Even with that qualification, the overall pattern is clear enough to matter: model rankings can change quickly when new entrants combine competitive capabilities with aggressive pricing, broad availability and an easy developer experience.

What the shift means for AI infrastructure

The week’s data points to a more contestable inference market. Premium U.S. proprietary models are no longer the only systems capable of attracting large international developer workloads through a shared routing layer. Chinese models can gain traffic by combining:

• low or temporary promotional pricing;
• free access that encourages experimentation;
• very long context windows;
• multimodal input;
• open-weight distribution; and
• perceived strength in coding and agentic tasks.

That does not prove that Chinese models have overtaken every competitor in capability, nor does it show that all of the traffic came from outside China. It does show that distribution and unit economics can change usage patterns rapidly. For teams choosing models, the relevant comparison is increasingly not “which model is best?” but “which model delivers the required result at acceptable cost, latency and risk?”

The August 24–30 snapshot therefore matters less as a permanent leaderboard than as evidence of how quickly inference demand can move. If the pattern persists after promotional access ends, it would provide stronger evidence that efficient Chinese models are winning durable developer adoption rather than simply benefiting from a launch-week spike.