Mysterious Hy3 LLM tops OpenRouter rankings: Why does Tencent’s open source model surpass Claude?

Core conclusion

In May 2026, a mysterious model named Hy3 quietly topped the OpenRouter AI Model Rankings, surpassing Claude Opus 4.7 and DeepSeek V4 Flash in terms of token consumption. What’s even more surprising is that Hy3 comes from Tencent’s open source warehouse and its benchmark test results are not outstanding, but the usage of paid users has continued to grow for more than three weeks.

Three key findings:

• 98% of token consumption is input token: Among the actual costs of LLM API calls, the proportion of output tokens is very small, suggesting that caching has become the core of pricing.
• The effective cost of DeepSeek V4 Flash is only $0.018/1M tokens - nearly half cheaper than Hy3's $0.034, provided you choose the right service provider
• Hy3 may be a single large application behind it: The data shows that it is not used by multiple individual customers, but an undisclosed data processing application that uses it as the main model

Event background: Who is Hy3?

On May 26, 2026, BuzzFeed Senior Data Scientist Max Woolf published an in-depth analysis article, revealing a strange phenomenon: On the OpenRouter rankings, two models that had no sense of existence before—Hy3 preview and DeepSeek Flash V4—are beating Claude Opus by more than 50%.

Hy3 is Tencent’s open source large language model, released on Hugging Face. But the strange thing is:

• Its benchmark results are not pretty**, even lower than other Chinese open source models
• The only result from googling it is Tencent's own launch announcement
• Searching for Hy3 on HN returned only one unrelated post
• On Reddit, the discussion is more about "open source weight releases" than actual usage experience

But the data doesn't lie - since the paid version was launched on May 8, Hy3's usage on OpenRouter has continued to grow and has been running stably for three weeks.

LLM Economics: The True Costs 99% of People Ignore

The most valuable part of Max Woolf's analysis is not about Hy3 itself, but his thorough breakdown of the economics of the LLM API.

Key Figures:

What does this mean? **LLM's "price tag" has been seriously distorted. **

In actual operation, since 98% of calls are input tokens, and input tokens are highly cacheable, the effective price of DeepSeek V4 Flash's direct connection from DeepSeek is only $0.018/1M tokens - 47% cheaper than Hy3's $0.034/1M.

Why is Hy3 so popular? Clues given by data

Woolf ruled out several possibilities:

Not automatic App/SDK switching: The top 5 apps combined account for less than 1% of total Hy3 usage Not a free strategy: The paid version will be launched on May 8, and the data is an ongoing voluntary payment behavior It’s not that quality beats the big manufacturers: Woolf’s actual measurement confirms that the quality of Hy3 is comparable to other Chinese models, far inferior to Claude Opus 4.7

The only clue is: Hy3 only has one service provider - SiliconFlow in Singapore. SiliconFlow had little usage before Hy3 came online. Woolf's guess is that a large data processing application (non-coding Agent class) is using Hy3 as a backend model, but the application does not disclose this choice.

"The advantage of OpenRouter is that the threshold for switching models and service providers is very low. I am not surprised that DeepSeek V4 Flash will reach the top in a few weeks - once everyone settles the score."

Practical inspiration for AI practitioners

1. Don’t just look at the list price, look at the effective price

When you select a model in Cursor, Codex, Claude Code, neither the subscription fee nor the list price equals the true expenditure**. DeepSeek V4 Flash The effective cost of direct connection from DeepSeek (2% cache read) means that your actual token cost may be only one-tenth of the listed price.

2. Caching strategy changes everything

98% input token ratio + ultra-low cache read cost = More and more AI workflows will call the same model repeatedly in the context instead of switching frequently. This has a direct impact on your Agent architecture design - prioritize cache-friendly workflow patterns.

3. China model vs data compliance

Woolf candidly pointed out: DeepSeek is a Chinese company, and some people may not be willing to hand over payment information or LLM input data to a Chinese company with prompt training set to true. For scenarios with compliance requirements, SiliconFlow's Singapore node may be a compromise.

Why this matter deserves attention

Hy3 topped the OpenRouter rankings. On the surface, it is a mysterious Tencent model that unexpectedly became popular. In fact, it is a turning signal of LLM economics:

1. Model homogeneity is accelerating - when users vote with their feet, price (especially effective price) becomes the decisive factor rather than quality differences
2. Caching optimization is the focus of the next round of competition - DeepSeek’s innovation in KV caching allows it to provide services at extremely low cost, and other manufacturers must follow suit
3. OpenRouter’s competitive model is becoming a pricing game - whichever model can provide the lowest effective cost will get the most token consumption

Suggestions for next action

• If you are using Claude Code or Codex, try switching part of your work to DeepSeek V4 Flash (directly connected to DeepSeek API) and compare the effective cost and output quality
• Set up cache hit rate monitoring: If your workflow cache hit rate is less than 60%, the architecture may need to be adjusted.
• Focus on the effective pricing of LLM rather than the list price - Woolf's article proves that the list price is distorted in the face of caching

Tool entry

The text naturally involves tool brand names such as OpenRouter, DeepSeek, Claude, Gemini, ChatGPT, Claude Code, Codex, etc.

Reference sources

• Max Woolf: The mysterious Hy3 LLM is topping OpenRouter Model Rankings
• HN 讨论：84 分，61 条评论
• OpenRouter AI Model Rankings

Internal link guidance

• Want to systematically understand model selection? Watch: AI 编程 Agent 技术选型：语言、模型、成本三维决策框架
• Actual combat effect: 用 DeepSeek V4 + Claude Code 搭微 SaaS 矩阵，月入 $8,500
• Money-saving solution: 如何用 DeepSeek V4 替代 Anthropic 跑 Claude Code：5 分钟省 90% API 费用