The OpenRouter whale

Between late April and mid-May, one model sat at the top of OpenRouter's usage rankings: Tencent's Hy3 preview, which pushed roughly 7.7 trillion tokens through the routing service in 19 days — a 54% lead over the second-place model. That figure isn't in dispute; it turns up in Tencent's own quarterly results. What's strange is everything around it. Hy3 wasn't the model people were arguing about, benchmarking, or recommending. It just quietly became the busiest endpoint on the platform.

Overall, I still don't understand the popularity of Hy3 preview on OpenRouter. — Max Woolf

Max Woolf, a data scientist who does this kind of forensics for sport, went looking for the reason and couldn't find one. It isn't quality — he rates Hy3 below the leading models from Anthropic and OpenAI. It isn't headline price either: once you account for the sole provider's cache-read fees, its effective cost works out to nearly double DeepSeek's cheapest model, so anyone optimizing purely for cost per token had a better option sitting right there. And the traffic has a peculiar fingerprint — by Woolf's read of the platform's stats, 98% of it is input, 2% output, the shape of a bulk data-processing pipeline chewing through documents, not a chatbot or coding tool. The apps OpenRouter names as the top users account for under 1% of it.

That's the quiet, useful point buried in a mystery nobody solved. OpenRouter's rankings are widely cited as a proxy for which models developers are actually adopting — but the board sorts by raw token volume and doesn't disclose how many distinct users are behind it. A single customer running one enormous batch job is indistinguishable from a groundswell of real adoption. Woolf's best guess is that some large, unaffiliated operation is using Hy3 as cheap bulk-processing plumbing, but he's careful to flag it as a guess. The honest conclusion is that the number-one model on the most-cited LLM leaderboard got there for reasons its own analysts can't reconstruct — which is worth remembering the next time a ranking chart is offered as evidence of what the field is choosing.