A little known AI model is beating Claude on one of the AI world's most watched usage charts, and almost nobody can explain why. It is called Hy3. On OpenRouter's public rankings, it is handling more traffic than famous names like Claude, often by more than 50 percent.
OpenRouter lets apps reach many AI models through one connection, so its leaderboard is a rough map of what people actually use. Usually the big brands win that kind of popularity contest. Right now, the chart is being won by something most people in the field had barely heard of.
A public leaderboard mystery
The puzzle was laid out by Max Woolf, a data scientist who writes about AI, on minimaxir.com. He checked the OpenRouter rankings in late May and saw two new models pushing past Claude. One, DeepSeek V4 Flash, made sense because it is a cheap model from a known Chinese lab.
The other one did not. Hy3 had almost no public chatter, no useful Hacker News trail, and only a couple of Reddit threads. Weird, right?
So what is Hy3?
Hy3 is an open weight model from Tencent, the Chinese giant behind WeChat and a huge games business. Open weight means the finished model files are public enough for others to download and run, instead of being locked behind one company service.
The model uses a mixture of experts setup, where many small expert networks sit inside one system and only some activate for each request. That can keep running costs down. The catch is that Tencent's own scores do not make Hy3 look special. They put it behind several other Chinese open models.
The numbers get stranger
Woolf tested Hy3 and found quality closer to middling Chinese models than to top systems like Claude Opus or GPT 5.5. So, no hidden gem. He also checked whether one popular app had quietly switched to Hy3 and dragged the chart upward, the way a coding tool once boosted a free model. That did not fit either, because the top five Hy3 apps added up to less than 1 percent of its traffic.
That leaves pricing, and even that gets messy. Modern AI assistants reread the whole conversation every time you send a message, so input tokens often dominate the bill. In agent-style tools doing long tasks, Woolf found input can hit 98 percent of cost. Prompt caching cuts that by reusing already processed text instead of charging full price again.
The cheap model problem
Here the twist arrives. Hy3 advertises a low input price, but once caching is counted its real cost works out to around $0.034 per million input tokens, while DeepSeek V4 Flash, run by DeepSeek itself, lands around $0.018 for the same amount. Basically half. That is ridiculous if the whole story is supposed to be "cheap model wins leaderboard."
So Hy3 may not be a mystery model everyone secretly loves. Actually, Woolf's best guess is narrower, one large company, not Tencent, may be running Hy3 behind some data-heavy product. That is still only a guess. The one sure thing is better, a model nobody can name is winning a race nobody saw it enter.
