Tencent's Hy3 Preview tops OpenRouter rankings with cheap $0.066/M tokens

Tencent's Hy3 preview, an open-source Mixture-of-Experts (MoE) model released in April 2026, has taken the OpenRouter platform by storm. With 295 billion total parameters but only 21 billion active per inference, it achieves a 14x efficiency gain. The model supports a 250,000-token context window and offers three inference modes balancing latency and depth. Priced at just $0.066 per million tokens, it undercuts DeepSeek V4 Flash ($0.10) and most competitors, making it a darling for budget-conscious developers.

From April 27 to May 11, Hy3 preview held the #1 spot on OpenRouter's weekly usage rankings, processing 3.66 trillion tokens in one week (298% week-over-week growth). It still sits at #2 even after the free trial ended, now behind DeepSeek V4 Flash. Tencent reported a 15.4% market share across all providers, with heavy use in coding agents like Hermes Agent and Claude Code. Its applications span academia, finance, health, legal, marketing, and programming.

However, benchmark results tell a different story. BuzzFeed's data scientist Max Wolf found Hy3 preview's performance on par with other Chinese models but far behind Claude Opus 4.7 and GPT-5.5. The model excels in software engineering (SWE), terminal tasks, and web browsing, but struggles in general reasoning. Wolf also warned that cached prompts could make DeepSeek V4 Flash cheaper, potentially shifting usage trends. Hacker News analysts noted that OpenRouter rankings favor high-token users, which may exaggerate Hy3's popularity.

Despite quality concerns, Hy3 preview's low cost and efficiency have reshaped the competitive landscape. Its success signals a market shift toward practical, affordable AI for mass applications, challenging premium models to justify their higher prices.