A major CNBC investigation published July 7, drawing on API routing data from OpenRouter and Vercel, has confirmed that Chinese AI models now account for between 30% and 46% of all weekly API token usage flowing through US developer platforms — a dramatic rise from approximately 11% over the prior 12 months and as low as 4.5% in the first half of 2025. The data represents the clearest quantified evidence yet that model-agnostic routing infrastructure has created a competitive opening for Chinese AI providers that is actively being exploited at significant scale by US enterprise developers.
The Price Gap Is the Core Driver
The mechanics driving Chinese model adoption are not primarily about capability parity — though capability gaps have narrowed substantially — but about price. Justin Summerville at OpenRouter quantified the differential bluntly: open-source Chinese models are currently 60 to 90 percent cheaper per token than leading Anthropic and OpenAI models for comparable tasks. At that magnitude of price difference, developers operating at scale have both commercial incentive and technical infrastructure to route workloads to cheaper alternatives.
Z.ai’s GLM-5.2 has been the fastest-moving data point in this story. Vercel tracking showed GLM-5.2 achieving a 27x increase in daily token volume and an 80x increase in customer count in its first full week after launch. The model scored 62.1% on SWE-bench Pro — a benchmark that measures software engineering capability — placing it in the vicinity of Western frontier models, while pricing at $1.40 per million input tokens and $4.40 per million output tokens compared to Claude Sonnet 5 at $2/$10 and GPT-5.5 at $2.50/$15.
What Multi-Model Routing Changes
The structural shift enabling this adoption is not the existence of competitive Chinese models — that has been the case for over a year — but the widespread deployment of model-agnostic routing infrastructure. Platforms like OpenRouter and Vercel’s AI SDK allow developers to switch between models with minimal code changes, treating AI providers as fungible compute suppliers rather than sticky platform choices. This infrastructure makes price-driven switching decisions that previously required significant engineering investment trivially easy to execute.
For Western frontier labs whose competitive moats have historically rested on capability leadership and developer ecosystem lock-in, the commoditisation of model access represents a structural challenge that model performance alone cannot fully address. The 80x customer growth for GLM-5.2 in a single week demonstrates that the switching friction which protected premium-priced models has declined to near-zero for cost-sensitive developer segments.
Policy Implications
US lawmakers have responded to the CNBC investigation by announcing formal probes into the growing use of Chinese AI models in domestic enterprise environments, raising questions about data handling, model provenance, and national security implications of routing sensitive workloads through infrastructure connected to Chinese AI providers. The investigation sits at the intersection of the AI competitiveness debate and the broader technology decoupling tensions that have characterised US-China trade policy across multiple technology sectors.
