Chinese AI Models Now Top 30% of OpenRouter Traffic as Cost Gap Hits 90%
Agentic AIChinese AI models have crossed the 30% mark on OpenRouter every week since February — and the real story isn’t market share, it’s the 60 to 90 percent cost advantage that’s driving the shift.
The Number
According to OpenRouter employee Justin Summerville, Chinese open-source models now regularly exceed 30% of weekly API traffic on the platform, up from an average of 11% a year ago. The share has hit 46% in some weeks. The Decoder’s analysis of the OpenRouter data identifies two catalysts: a dramatic cost differential and performance parity on the workloads that actually run at scale.
How the Math Works
The cost structure is the engine. Chinese open-source models run 60 to 90 percent cheaper than US counterparts on comparable API calls. Startup Lindy provides the clearest signal: CEO Flo Crivello confirmed the company shifted all inference traffic from Anthropic’s Claude to DeepSeek, and that the switch “saves millions.”
DeepSeek and Z.ai are the primary Chinese providers gaining share with US teams. According to Brookings Institution’s Kyle Chan, Chinese models trail US counterparts by roughly 6 to 9 months on capability benchmarks — consistent with a May 2026 government report that found an approximately 8-month gap across cybersecurity, software development, math, and reasoning. For the workloads that actually consume most API tokens — classification, summarization, document processing, structured extraction — that gap is irrelevant.
What It Signals
OpenRouter traffic is a leading indicator of developer preference, not consumer sentiment. These are teams choosing their inference layer, not individuals picking a chat interface. When a company shifts “all traffic” from a US frontier model to a Chinese open-source alternative, the switching cost is real — meaning the savings are large enough to justify it.
The second implication is pricing pressure. US providers have been cutting prices, but Chinese models have been cutting faster. A 60 to 90 percent cost delta is not a rounding error; it’s a structural advantage that compounds across token-intensive agentic workflows. Teams running millions of inference calls per day face materially different economics than teams running thousands.
The So What
For teams running high-volume, non-frontier workloads, the question is no longer whether to consider Chinese models — it’s what the actual risk calculus looks like for your use case. The capability gap is smaller than the cost gap. Compliance and data-residency risks are real but bounded and highly dependent on workload type. Teams building production systems at scale will run this cost-benefit analysis in the next two quarters, and a significant portion will move at least some traffic. The 30% figure will keep climbing.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →