The counter on OpenRouter’s dashboard keeps climbing. The service, which routes traffic from developers to hundreds of AI models, recorded 28.9 trillion tokens of usage in the week ending May 24 — a 7.4% increase over the previous week and the fifth consecutive weekly gain, according to data published by the company. Behind the aggregate number is a shift that would have been unthinkable two years ago: the two most-used models in the world are now Chinese.
DeepSeek-V4-Flash topped the weekly rankings with 3.43 trillion tokens, followed by Hy3 preview with 3.07 trillion, according to OpenRouter’s data. The two Chinese models accounted for more than a fifth of all traffic measured by the service, ahead of the flagship models from OpenAI, Anthropic and Google. The rankings reflect a structural change in the AI market, where low-cost Chinese models have become the default choice for a growing share of developers.
The numbers need context. OpenRouter measures only the traffic that flows through its own platform — a large and representative slice of API usage, but not the whole market. Consumer apps like ChatGPT, which handle enormous volumes of their own traffic, are not fully captured in the figures. Even with those caveats, analysts said, the trend is unmistakable: open and low-priced models are taking share, and Chinese labs are the ones supplying them.
DeepSeek’s rise is the industry’s most disruptive story of the past two years. The company’s models matched frontier performance at a fraction of the cost, upending assumptions about how much intelligence should cost, and its pricing forced every major lab to respond with cheaper tiers and faster models. V4-Flash, the latest iteration, extends the pattern: it is fast, cheap and good enough for the bulk of everyday AI work — the classification, extraction and generation tasks that make up most real-world traffic.
The weekly growth is significant for another reason. Five consecutive weeks of rising global token consumption suggests that AI usage is still expanding rapidly, even as the industry’s attention has shifted from growth to profitability. For the companies building data centers and buying chips, the numbers support the case for continued investment: demand, measured in actual tokens consumed, keeps compounding. For the skeptics who argued that AI usage would plateau, the OpenRouter data offers little comfort.
The pattern in the rankings also says something about how AI is being used. The most-trafficked models are not necessarily the most powerful; they are the ones that are fast, cheap and reliable for high-volume work. Developers route their routine tasks to the cheapest adequate model and reserve the most expensive frontier systems for the jobs that need them. That cost-conscious behavior, multiplied across thousands of teams, has produced the current rankings — and it is why the cheap Chinese models lead.
The token numbers carry economic weight. Every trillion tokens represents real spending on computing, and the growth in usage is translating into revenue for the cloud providers and chipmakers that supply the underlying infrastructure. The economics have shifted: where the industry once monetized scarcity, it now monetizes volume, and the winners will be the companies that can serve the most tokens at the lowest cost. The Chinese labs’ advantage in that game is their pricing discipline.
The trend has geopolitical dimensions as well. Two years ago, the conventional wisdom held that the U.S. had an unassailable lead in AI; today, the most-used models in the world are Chinese, and the gap in capability has narrowed to the point where cost, not quality, is the deciding factor for many buyers. The shift has not gone unnoticed in Washington, where policymakers have debated export controls and domestic investment in response to China’s rapid progress.
The weekly figures also obscure enormous variation beneath the aggregate. Token volumes are dominated by a long tail of applications — customer service bots, document processing pipelines, coding assistants, translation services — that each consume modest amounts individually but together account for the bulk of usage. The stability of the rankings, with the same models holding the lead week after week, suggests the distribution is settling into a pattern: a few workhorse models carry the world’s routine AI traffic, while frontier systems serve a smaller, more specialized demand.
For the companies at the top of the rankings, the position brings a new kind of pressure. Being the most-used model in the world invites scrutiny — of safety, of data handling, of the consequences of putting a Chinese model at the center of global software. The labs that built the models have shown they can compete on price and performance; whether they can manage the responsibilities of scale is the next test. For now, the tokens keep flowing, and the counter keeps climbing.
