Tencent's Hy3 LLM Mysteriously Dominates OpenRouter Rankings Despite Lower Quality Benchmarks
Key Takeaways
- ▸Tencent's Hy3 has unexpectedly become the most-used LLM on OpenRouter despite underperforming compared to Claude and GPT in standard benchmarks
- ▸The model's popularity likely stems from its low cost ($0.066/M tokens) combined with specialized use case efficiency, not superior quality
- ▸The 98% input, 2% output token breakdown suggests Hy3 is optimized for lightweight, input-heavy inference tasks like batch processing or analysis
Summary
An open-source language model from Chinese tech giant Tencent called Hy3 has unexpectedly become the most-used LLM on OpenRouter, beating OpenAI's GPT and Anthropic's Claude by over 50% in token usage, according to analysis by data analyst minimaxir. Despite the impressive ranking, benchmark results show Hy3 performs significantly worse than leading models like Claude Opus 4.7 and GPT 5.5, raising questions about what's driving its unexpected popularity.
The surge began on May 8, 2026, when Hy3 switched from a free tier to paid access on OpenRouter at just $0.066 per million input tokens—cheaper than competitors like DeepSeek V4 Flash at $0.10/M tokens. However, the pricing advantage alone doesn't fully explain why users would favor a lower-performing model at scale. Analysis reveals an unusual usage pattern: 98% of Hy3 tokens are input tokens with minimal output, suggesting the model may be serving specific, compute-light use cases rather than general-purpose tasks.
The mystery deepens when examining adoption patterns. Unlike other popular models that see spikes when major applications switch defaults, Hy3's adoption appears organic and steady since the paid launch. OpenRouter lists only one provider for Hy3 compared to 13 for DeepSeek V4 Flash, limiting distribution—yet demand remains robust. This phenomenon underscores an emerging trend: cost optimization and specialized use case efficiency may now outweigh raw performance benchmarks in driving LLM adoption decisions.
- The rise of Hy3 signals a market shift where cost efficiency and task-specific suitability may be gaining importance over raw model performance
Editorial Opinion
Hy3's unexpected dominance reveals a growing divergence in how users evaluate LLMs: raw benchmarks and brand recognition no longer tell the complete story. The model's popularity despite lower quality benchmarks suggests that practical considerations—cost, latency, and task-specific suitability—may be reshaping the competitive landscape more profoundly than vendors realize. If this trend continues, we could see a future where specialized, cost-efficient models dominate everyday use cases while only the most demanding applications rely on premium, large-scale models.
