News · 2026-08-13

Chinese models passed American ones in OpenRouter traffic in June

Chinese models passed American ones in token share on OpenRouter in early June 2026, according to OpenRouter's own published analysis. The swing was driven mostly by a single model: DeepSeek V4 Flash captured about 70 percent of DeepSeek's agentic token flow by the end of May, while DeepSeek as a whole climbed from around 5 percent of weekly platform tokens in February and March to nearly 20 percent by the start of June. A separate claim circulating this week -- that open-weight usage had "fallen below 50 percent" -- does not survive contact with the same source.

Key facts

• OpenRouter dates the crossover where Chinese models surpassed American ones in token share to early June 2026.
• DeepSeek started 2026 just under 10 percent of weekly OpenRouter tokens, dropped to roughly 5 percent in February and March, and recovered to nearly 20 percent by early June.
• V4 Flash took about 70 percent of DeepSeek's agentic token flow by the end of May, with Xiaomi, MiniMax, Tencent, and Qwen families also gaining as Google and OpenAI plateaued on the platform.
• Primary sources: OpenRouter, DeepSeek V4 Is Earning Agentic Token Share (June 30, 2026) and the State of AI 2025 usage study.

The correction is worth making carefully, because the wrong version of this story has now run twice in opposite directions. Earlier this summer the claim was that open models carried most of the traffic. This week the claim was that open-model usage had fallen below half. Both are wrong for the same reason: OpenRouter's own numbers never had open weights above half at any point. Its State of AI report puts open-weight models at roughly one third of platform usage by late 2025, with proprietary models serving the majority throughout, and no single open model exceeding about a quarter of the open-model slice.

What did cross over is a different axis. The line that moved is national, not licence-based: Chinese-origin models overtaking American-origin ones in token share. Those two axes correlate strongly, because most prominent Chinese releases ship open weights, which is exactly how they get conflated. But they are not the same measurement, and the difference determines what conclusion you are entitled to draw. "Open beat closed" is a claim about business models. "Chinese labs beat American ones on this platform" is a claim about who is shipping the models developers reach for, and only the second is supported.

The mechanism behind the shift is narrower still, and it is agentic workloads. V4 Flash is cheap, fast, and good enough at tool calling to run inside long agent loops where token consumption is enormous -- a workload that inflates token share far faster than it inflates user count. One popular agent burns more tokens in an afternoon than a thousand chat users do in a week. That is how a single model takes 70 percent of a lab's traffic in a month, and it is why token-share charts move so violently.

The strongest available counter-argument comes from the sell side. On Morgan Stanley's Thoughts on the Market, analysts note that while OpenRouter data suggests open-source models will take a lot of tokens, "the vast majority is still the proprietary models" on a dollar-weighted basis. Both statements can be true at once, and they usually are: cheap models win token counts precisely because they are cheap, which is the same reason they lose revenue counts. Commentary on Hacker News has pressed the same point from the technical side, arguing that token share on one aggregator is not unique users, not enterprise deployments, and not evidence that open weights have caught the frontier on high-value coding work.

Provenance is also worth stating plainly. Hugging Face's chief executive Clement Delangue has argued repeatedly in interviews that China leads on open models and may reach the frontier soon. A specific CNBC quotation to that effect circulated this week; we could not match it to an accessible primary transcript, so we are not reproducing it. The underlying position is well documented in Delangue's other public statements, and it should be read with the obvious interest attached: Hugging Face is the distribution layer for the open ecosystem whose ascendancy it is describing. Hugging Face's own trending board today is heavily weighted toward Chinese open-weight families, which is board-level evidence for the direction and not for the magnitude.

The near-term complication is that the cheap end just got more expensive. DeepSeek is replacing flat API pricing with peak and off-peak rates on August 17, raising every hosted rate and hitting cached input on its Pro model hardest -- which is exactly the agent workload that produced this token-share story in the first place. If hosted DeepSeek gets pricier during working hours, some of that traffic moves to self-hosted weights, which no aggregator chart captures at all, and some moves to whichever provider is cheapest that week. Token-share leadership built on being the cheapest option is leadership with a short lease.

Primary source, verified: read the paper →

Ground Truth. (2026, August 13). Chinese models passed American ones in OpenRouter traffic in June. Ground Truth. https://groundtruth.day/news/chinese-models-passed-american-ones-in-openrouter-traffic-in-june.html

@misc{groundtruth:chinese-models-passed-american-ones-in-openrouter-traffic-in-june,
  title  = {Chinese models passed American ones in OpenRouter traffic in June},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/chinese-models-passed-american-ones-in-openrouter-traffic-in-june.html}
}

Topics: open-weights · china · deepseek · market-share · openrouter · industry