Chinese AI Models Led OpenRouter Usage for 15 Straight Weeks

OpenRouter’s August 3–9, 2026 usage snapshot showed Chinese-developed large language models extending their lead over U.S. models for a 15th consecutive week. Across the models tracked by the routing platform, Chinese models generated a reported 34.25 trillion tokens, compared with 9.17 trillion for U.S. models, while total weekly volume reached 69 trillion tokens—up 21.48% from the prior week. 313

That is a significant signal about the preferences of developers using a multi-model API marketplace. It is not, however, proof that Chinese models account for most LLM usage worldwide: OpenRouter does not include traffic sent directly through providers’ own products and APIs. 30

The August 3–9 OpenRouter numbers

The reported figures show a wide gap in routed usage:

• Total tracked token volume: 69 trillion, up 21.48% week over week. 313
• Chinese-developed models: 34.25 trillion tokens, up 21.76%. 313
• U.S.-developed models: 9.17 trillion tokens, despite a reported 109.36% weekly increase. 313
• Leadership streak: Chinese models ranked ahead of U.S. models for the 15th consecutive week. 913

Chinese models therefore represented roughly half of the tracked total, while the U.S. group remained well behind despite growing faster from a smaller base. The comparison is best read as a measure of model mix on OpenRouter, rather than as a direct measurement of national AI market share.

Chinese models swept the top four positions

The clearest sign of concentration was the individual leaderboard. Reports based on the OpenRouter data placed four Chinese models at the top:

1. DeepSeek-V4-Flash-0731: 8.83 trillion tokens, up 570% week over week
2. Tencent Hy3: 8.05 trillion tokens, up 67%
3. DeepSeek-V4-Flash-0423: 5.88 trillion tokens, down 19%
4. Xiaomi MiMo-V2.5: 5.39 trillion tokens, down 14% 91013

The figures show that the lead was not produced by a single breakout model alone. DeepSeek’s official 0731 release took first place, but Tencent and Xiaomi also held major positions, while the earlier DeepSeek preview remained in the top four. That breadth makes the result more meaningful than a one-week spike from one provider.

Some later reports give different token totals for comparable OpenRouter windows. The numbers above are the figures repeated in the August 3–9 reports used for this analysis; they should be treated as reported platform data, not as independently audited global measurements. 45

Why DeepSeek-V4-Flash-0731 mattered

DeepSeek-V4-Flash-0731 was released into public API beta on July 31. Available reporting describes it as a retrained version of the earlier V4-Flash architecture rather than a larger replacement: the model retained the same 284-billion-parameter mixture-of-experts design, with 13 billion active parameters per token. 3940

The reported improvement came from redoing post-training, the stage that shapes how a model uses its underlying capabilities. Coverage of the release linked the new checkpoint to stronger agentic, coding, and math-oriented evaluation results, while emphasizing that the architecture and parameter count had not changed. 343639

That distinction matters for developers. The OpenRouter ranking suggests that model adoption can move quickly when an existing architecture receives a useful post-training update and is available through a low-cost, high-capacity API. The data does not prove that benchmark gains alone caused DeepSeek’s 570% increase, but the timing of the July 31 release and the subsequent usage jump make the release an important context for the result. 31340

Claims that the model averaged exactly $0.03 per benchmark task, cost one-hundredth as much as a named Anthropic model, or was definitively the world’s lowest-cost major model are not established by the evidence reviewed here. API prices also vary by provider and routing arrangement; available listings show different rates across DeepSeek’s first-party API and OpenRouter. 204046

OpenAI’s Luna surge shows price still mattered

The ranking did not mean U.S. models were absent. OpenAI’s GPT-5.6 Luna was reported in fifth place at 4.43 trillion tokens, with usage up 128% during the week. 30

OpenAI had officially cut Luna’s API price by 80% on July 30, bringing its listed price to $0.20 per million input tokens and $1.20 per million output tokens. 20 The timing makes lower pricing a plausible contributor to Luna’s increase, although the available data does not establish a single cause.

This is an important counterpoint to the broader China-versus-U.S. framing. Developers appear responsive to price reductions from any major provider. Chinese models’ lead cannot be explained simply by the existence of a cheaper alternative, but pricing remains one of the competitive levers shaping model selection.

What the ranking actually reveals

The strongest conclusion is narrower—and more useful—than “Chinese AI has won globally.” On OpenRouter’s developer-oriented routing marketplace, Chinese models were winning usage at scale while occupying the leading positions across several products.

That pattern is consistent with a competitive advantage built from multiple factors:

• Cost: low API prices can make large-volume workloads economically viable.
• Throughput and capacity: practical deployment depends on serving more requests reliably, not only on benchmark scores.
• Task performance: coding, tool use, agent workflows, and math can influence which models developers keep in production.
• Availability and flexibility: models exposed through a common routing layer are easier to test and switch between.
• Observed adoption: token volume reflects what developers actually routed during the week, even if it does not explain why they chose each model.

The top-four sweep is therefore stronger evidence of real deployment preference than a benchmark table alone. At the same time, token counts do not reveal the quality of every request, the identity of users, profitability, or how much usage came from experiments versus durable production workloads.

What the data does not prove

OpenRouter’s numbers should not be treated as a complete global leaderboard. Direct use of ChatGPT, provider-native APIs, enterprise contracts, and other routing platforms falls outside the snapshot described here. 30

The sources reviewed also do not independently confirm the full set of claims about MiniMax M3 and Stepfun Step 3.7 Flash leaving the top 10, Gemini 3.6 Flash reaching exactly 2.33 trillion tokens with 446% growth, or a precise benchmark-task cost comparison for DeepSeek. Those details should be verified against a primary OpenRouter dataset or model-provider documentation before being presented as settled facts.

Bottom line

The August 3–9 data showed Chinese models leading OpenRouter usage for 15 straight weeks, with 34.25 trillion tokens versus 9.17 trillion for U.S. models and all four top positions held by Chinese systems. 313

The most defensible interpretation is that Chinese models had become exceptionally competitive in the developer channel represented by OpenRouter—not just because they were inexpensive, but because price, serving capacity, recent post-training improvements, and usable performance were aligning at the same time. The caveat is decisive: OpenRouter is a valuable usage signal, not a definitive measure of global LLM consumption.