US developers are routing a growing share of their AI workloads to Chinese open-weight models, according to gateway-traffic data reported by CNBC on July 7. The share of tokens that US companies send to Chinese models through the OpenRouter gateway has stayed above 30% every week since early February, and has spiked as high as 46% — up from an average of just 11% across the prior twelve months.

The trigger: near-frontier scores at a fraction of the price

The shift tracks the release of Z.ai's GLM-5.2, an open-weight model that landed within about a percentage point of Anthropic's Claude Opus 4.8 on FrontierSWE, a closely watched long-horizon agentic-coding benchmark, while costing roughly a fifth as much to run. Z.ai prices GLM-5.2 at $1.40 per million input tokens and $4.40 per million output tokens, versus Anthropic's $10/$50 per million for Claude and OpenAI's roughly $5/$30 for GPT-5.5. Vercel, which tracks model usage across its infrastructure, says GLM-5.2 saw the fastest adoption of any model it monitored in 2026: daily token volume grew about 27x and the number of customers using it grew about 80x in its first full week after launch. DeepSeek is also gaining share — AI startup Lindy said it moved 100% of its production traffic off Claude models onto DeepSeek in June.

Price, not just capability, is now the deciding factor

"Price is doing the work here," Vercel's head of agentic infrastructure told CNBC, describing teams that once defaulted to the strongest available model now routing routine tasks to "the cheapest one that's good enough" — a trade Chinese open-weight models are increasingly winning as the performance gap narrows. Broadly, researchers estimate open-weight Chinese models now run 60-90% cheaper than comparable US frontier models.

Adoption still has real limits

Despite the token-share gains, analysts caution the shift is concentrated in cost-sensitive, non-regulated workloads. Political scrutiny of Chinese AI providers and data-security concerns remain major barriers to deeper adoption in the US, particularly in finance, healthcare and other regulated sectors, and most enterprises are pairing cheaper Chinese models with US frontier models rather than fully replacing them. Still, the sustained token-share numbers suggest the "good enough, much cheaper" calculus is now a durable part of how US teams choose which model handles which task — a dynamic that puts renewed pricing pressure on OpenAI and Anthropic heading into the rest of 2026.