A model called Hy3 preview is currently sitting at the top of OpenRouter's public model rankings, beating Claude by more than 50% in token usage. Two weeks ago, most engineers had never heard of it.
OpenRouter publishes weekly usage data that reflects real paying-user behavior across most major LLMs. That makes its rankings a reliable signal, not a lab-curated leaderboard. When an unfamiliar model jumps to the top, it deserves a close look.
Hy3 preview is an open-source release from Tencent. The model's own Hugging Face page includes benchmark results that are, in the words of researcher Max Woolf, "oddly honest" and unfavorable compared to other Chinese open-source models. Coding benchmark scores place it behind peers, not ahead. After hands-on testing, Woolf concludes the model quality is on par with similar Chinese open-source models and not close to top-tier models like Claude Opus 4.7 or GPT 5.5.
So why is it dominating usage?
Price is the first obvious answer. Hy3 preview is listed on OpenRouter at $0.066 per million input tokens. DeepSeek V4 Flash, the second-ranked model and a legitimately strong performer at low cost, is listed at $0.10 per million input tokens. That gap matters when you are running agents at scale.
But price alone doesn't explain it cleanly. Hy3 briefly had a free endpoint on OpenRouter in early May, which drove an initial spike. That free tier is now gone. The current rankings reflect paying users, not free-tier experiments. Usage is also spread across domains beyond coding agents, which rules out a single niche use case driving the numbers.
Woolf checked OpenRouter's deeper data and came away more confused, not less. No single clean explanation emerged. A cheaper price point, some residual momentum from the free period, and possible programmatic or automated usage at scale are all candidates. None fully account for the margin.
For builders, the practical read is this: OpenRouter's public rankings are worth watching as a cost-pressure signal, but high token volume does not equal high quality. Hy3's rise appears to be driven by price and possibly non-human traffic patterns, not by engineers choosing it for capability.
If you are selecting a model for a production coding agent or any quality-sensitive pipeline, verify independently. Run your own evals on your actual tasks. Do not let a usage ranking substitute for benchmark work on your specific workload. And if you are already using Hy3 because it appeared at the top of a list, now is a good time to double-check your output quality metrics.