Chinese models now carry 60% of OpenRouter traffic, and 58% of what US firms route

Chinese-developed models took all five top spots on OpenRouter by token volume in July, the first time that has happened, and now carry more than 60% of traffic on a platform that routes over 20 trillion tokens a week ^[1][3]. The more consequential figure is domestic: by mid-July, Chinese models accounted for a record 58% of the tokens processed by American firms on the platform ^[5].

OpenRouter is not a benchmark. Fortune describes it as a neutral routing platform that has become the closest thing the industry has to a Nielsen rating, which is to say it measures what people actually run rather than what they claim to prefer ^[21]. On that measure the reversal is fast: US models carried roughly 70% of the platform's traffic a year ago and about 30% now, a swing of some 40 percentage points ^[4][20]. At current volumes, Chinese models absorb upwards of 12 trillion tokens a week there ^[22].

The frontier claim is not in dispute. Fortune's own framing puts GPT-5.5, Claude Fable 5 and Gemini 3.x ahead on the hardest reasoning and long-horizon agent work ^[6]. The gap that is moving purchase orders is price. DeepSeek's V4-Pro is priced at roughly one-twelfth of GPT-5.5 at comparable benchmark performance, and DeepSeek V4 Flash lists at $0.14 per million input tokens against $5.00 for GPT-5.5, a ratio of about 36 to 1 ^[7][8][9]. OpenRouter's analysts put the general range for Chinese open models at 60% to 90% cheaper than leading American offerings ^[10].

Then there is the part that does not reverse quickly. Alibaba's Qwen family has passed a billion cumulative downloads and displaced Meta's Llama as the most downloaded open model family, while Llama itself has fallen below 1% of routed volume ^[11][12]. Tooling, fine-tunes and internal know-how accumulate around whatever engineers can download and self-host, and that stock is now being built on Chinese weights.

Two cautions before anyone rewrites a budget. Routed tokens are not dollars, and the source's own data proves it: Anthropic holds about 12% of OpenRouter's token share while capturing roughly half of total platform spending, roughly four times the platform's average revenue per token ^[13][14]. A premium lane priced for work that justifies it can lose volume share for a long time without losing revenue. Second, the cost base underneath the cheap lane is partly political. Fortune attributes Chinese efficiency to engineering around export-control scarcity through token efficiency, novel attention, mixture-of-expert designs and inference-aware architecture, plus state support that lowered the effective cost base, with Xiaomi cutting MiMo API prices by as much as 99% in May ^[15][16]. A 99% price cut is a market-share instrument, not a margin.

The exposed position is the middle: the closed model with no decisive capability edge, and the enterprise paying frontier prices for commodity work ^[17]. The 2024 pattern of signing one frontier API contract and routing everything through it is what created that exposure ^[18]. Fortune's prescription is hybrid routing, with claimed inference savings of 60% to 90% on the majority of workloads for firms already doing it ^[19].

What to watch: whether US list prices move toward the cheap lane rather than staying flat while volume drains, whether Anthropic's revenue share holds as its token share stays thin ^[13], and whether Qwen's download lead converts into production deployments inside American companies rather than experiments. Nothing in routed-token data tells you which workloads a regulated buyer is permitted to send offshore, and that constraint, not price, is what will cap the 58% figure if anything does ^[5].