Chinese AI Models Capture 48% of OpenRouter Traffic as U.S. Share Drops From 74% to 20% in One Year

My take: OpenRouter data from the past year documents a structural shift in how developers and platforms choose their AI models. In June 2025, US models — OpenAI, Google, and Anthropic combined — controlled 74% of token traffic on OpenRouter. By June 2026, that figure had dropped to 20%. At the same time, Chinese models climbed from 20% to 48% and now process around 18 trillion tokens per week, compared to 5.5 trillion for US models. The primary driver is price: models like Tencent Hy3 charge $0.063 per million input tokens, between 10 and 20 times less than equivalent US frontier models.

For those of us building with AI, this shift has a direct implication. Model routing — choosing which model runs which task based on price and capability — is no longer an advanced optimization: it is a baseline decision. Coding tasks, which account for more than 50% of all OpenRouter usage, are exactly the kind of work that can be delegated to lower-cost models without losing quality. The cost of not having that visibility is not marginal.

Do you know, on average, what each type of AI task costs you today, and are you choosing the right model for each one?