Hy3's Suspicious Rise Atop OpenRouter Rankings Defies Quality and Pricing Logic

A little-known Tencent model called Hy3 preview has quietly climbed to the top of OpenRouter’s usage rankings, outpacing Claude by more than 50% in token volume despite minimal public discussion, sparse documentation, and benchmarks that place it well behind frontier models like Claude Opus 4.7 and GPT 5.5. The model is served by a single provider, Singapore-based SiliconFlow, and its usage barely dipped when it transitioned from a free tier to paid on May 8, suggesting either genuine organic demand or expensive artificial inflation. App-driven default switching, the usual explanation for sudden ranking spikes, does not account for the pattern here.

The cheap-pricing theory also falls apart on closer inspection. While Hy3’s stated rate of $0.066 per million input tokens undercuts DeepSeek V4 Flash’s $0.10, effective pricing tells a different story once prompt caching is factored in. Input tokens now make up 98% of aggregate API traffic and are aggressively cached, and DeepSeek’s new KV caching approach gives it a 2% cache read cost when served directly, dropping its effective price to $0.018 per million tokens — roughly half of Hy3’s $0.034 effective rate through SiliconFlow’s 44% cache read cost.

The piece leaves the anomaly unresolved but flags it as a case study in how opaque LLM market dynamics have become, and how the geopolitics around Chinese-origin models complicates straightforward explanations of provider choice.