Tencent Hy4 Benchmark Results for Data Analysis

Nov 4, 2026: Tencent's new flagship model, tested on real-world data analysis.

At Plotly, we keep a close eye on model releases. Our agentic analytics products, Plotly Studio Desktop and Plotly Studio Embedded, and our open source libraries Plotly and Dash work with any model, and our customers look to us for recommendations.

On November 4, Tencent released Hy4, the full version of the Hy4 preview it open-sourced in August. Tencent says the release brings large gains over the preview in agentic work and long, real-world productivity tasks, and adds image understanding for the first time. It has a context window of about a million tokens. On OpenRouter it costs $0.42 per million input tokens, $1.25 per million output tokens and $0.021 per million cached input tokens.

We ran it through our data analytics exam. Hy4 passed 42 of 43 questions (98%) for $2.11 for the whole exam. Only GPT-6.1 Sol and GPT-6 Astra, with a perfect 43, scored higher. Grok 4.7 and Kimi K3.1 Pro also scored 42, but cost $7.85 and $13.47. That makes Hy4 the cheapest model on our leaderboard of 37 to score 42 or more, and it joins the cost-accuracy frontier between Particle-1 and GPT-6.1 Sol: no model we've tested is both cheaper and more accurate.

Hy4 vs. GPT-6.1 Sol and Grok 4.7

*1,000 people each asking 10 questions a day, 250 working days a year, at OpenRouter's prices when each model ran.

Hy4 aced five of the six kinds of question on the exam: the analytics questions, the ambiguous ones, the impossible requests, the data quality audits and the statistics traps. Its one miss was a hidden pattern. GPT-6.1 Sol scored one point higher for about 24% more money, and Grok 4.7 matched its score at almost four times the cost.

One run per model is a single sample. In our repeat runs of other models, scores moved by up to three points between runs, so treat a one-point gap as a tie.

Thorough, and slow

Hy4 is one of the slowest models we've tested. Its median question took 117 seconds, against 18 for GPT-6.1 Sol and 38 for Grok 4.7, and the whole exam took almost three hours. Among the models that scored 40 or more, only Fugu Ultra v2, Kimi K3.1 Pro and MiniMax M3.1 Flash were slower.

Most of that time goes into checking its work. Hy4 ran about 24 queries and charts per question, and on the questions where we hid patterns in the data it averaged 64, against a median of 26 across the board. It cross-checked counts against other tables, looked at trends month by month, and stated its assumptions and how confident it was. Those habits are what the hardest questions on the exam reward.

One question went badly wrong on time, though not on accuracy. Hy4 wrote a query that paired every row of one large table with every row of another, and that single query ran for almost an hour. The rest of its work on that question was sound, and its final answer was correct. Our harness doesn't put a time limit on queries, so in a product you'd want one.

Hy4 is also cheap to run despite all that work. It read about 22.7 million input tokens over the exam, but 91% of them came from the cache at $0.021 per million, so its input cost about $1.25 instead of about $9.50.

Where it got things right

• Statistics traps (5 of 5). Ten of the 37 models on our board, Hy4 included, aced these. Hy4 broke aggregates down before drawing conclusions and named the rule it used whenever it called something unusual.
• Data quality (3 of 3). It audited each table before using it and caught every problem we planted.
• Hidden patterns (4 of 5). Only five models have found all five.
• Analytics (20 of 20), ambiguous questions (5 of 5) and impossible requests (5 of 5).

Where it got things wrong

Hy4 missed one of the 43 questions. Asked whether any components were failing more often than they should, it found a real and large excess in one component type, measured it carefully over two years, and was candid about what the data could and couldn't show. But it never looked at the part of the data where we hid the pattern, and it ruled out a cause it hadn't actually checked.

Open weights

Tencent published the weights for the Hy4 preview under the Apache 2.0 license, and says it will release the weights for the full Hy4 in the coming days. For now Hy4 is available through Tencent Cloud and OpenRouter. Once the weights are out, Hy4 will tie Kimi K3.1 Pro, whose weights Moonshot released on November 2, as the most accurate open-weight model we've tested, at about a sixth of the cost per run. We wrote more about that race in Open Models vs. Closed Models in 2026.

Frequently asked questions

How good is Tencent Hy4 at data analysis?

On PlotlyDataAnalyticsBench it passed 42 of 43 questions, tied with Grok 4.7 and Kimi K3.1 Pro and one behind GPT-6.1 Sol and GPT-6 Astra. It aced the analytics, ambiguous, impossible, data quality and statistics questions, and found four of the five hidden patterns.

How much does Hy4 cost?

On OpenRouter, $0.42 per million input tokens, $1.25 per million output tokens and $0.021 per million cached input tokens. Our whole exam cost $2.11, the least of any model that scored 42 or more.

How fast is Hy4?

Slow. Its median question took 117 seconds, more than six times as long as GPT-6.1 Sol, because it checks its work with many more queries.

Is Hy4 open source?

Not yet. Tencent released the Hy4 preview's weights under Apache 2.0 in August, and says it will release the full Hy4's weights in the coming days.

About the benchmark

We run new models through the same exam as they're released, and we'll keep sharing what we find. We ran Hy4 through OpenRouter at its default settings, with the same harness and grading as every other model.

To see all 37 models on one chart and in a full results table, along with cost, speed and how the exam works, see Introducing DataAnalyticsBench.

The exam is private and we're sharing it with researchers. We're keeping it private so that the answer key doesn't make its way into the training set. For the same reason, the data is fabricated: the warehouse is a synthetic wind-energy fleet, built to look like a real operator's data, so its answers don't exist anywhere else.

If you'd like access, reach out to Chris Parmer at Plotly.