Hunyuan Hy4 Tops Chinese Models on Agent Arena; 12K Real Tasks Are More Worth Watching Than the Ranking

Arena.ai update on November 6: Tencent Hunyuan Hy4 ranks first among Chinese models in Agent Arena and 14th overall; across 12K+ real, long-horizon Agent sessions, it shows a net improvement of +5.4%, with a median cost of $0.57 per task. It is also the only model from a Chinese lab in the top five.

Agent Arena collects multi-step tasks that users hand to agents, including real operations such as tool calls, error correction, and command-line execution. Single-turn Q&A is not included in this scope. The samples come from users worldwide, not from questions the lab itself created.

Put the data from two days earlier alongside this. On November 4, Hy4 was 14th overall, with 8.5K+ sessions, a net improvement of +4.7%, and a cost of $0.55 per task. Arena.ai stated at the time: Hy4 is currently still proprietary, and only after the weights are released will it tie with Kimi K3.1 Pro for first among open models. This qualification is crucial. The leaderboard is saying "once the weights are released"; the weights have not been released yet.

At $0.55/task versus Kimi K3.1 Pro Max's $1.46/task, it is about 62% cheaper; the score gap between the two is within 1 percentage point. Its position on the Pareto frontier also comes from this combination of cost and score.

From preview to official version, Hy4 rose from 25th with +1.1% to 14th with +4.7%, an improvement of +3.6 percentage points. The sub-scores moved as well:

• Confirmed Success: 11th, +8.7%, previously 16th, +4.9%
• Praise vs Complaint: 18th, +3.8%, previously 24th, +1%
• Steerability: 14th, +5.7%, previously 27th, +0.21%
• Bash Reliability: 20th, +6.2%, previously 41st, -3.1%

Confirmed Success is the proportion of users who explicitly click 'yes, that worked'; Steerability looks at how the model responds to user corrections; Bash Reliability looks at the success rate of shell commands.

As the sample grew from 8.5K to 12K, the net improvement went from +4.7% to +5.4%, and the median cost went from $0.55 to $0.57. In the replies, geekzard said that the part that convinced him was that it still holds up as the sample gets larger. Adrian Castellar, meanwhile, pointed out that Hy4's +5.4% and Kimi K3.1 Pro's +5.3% have overlapping error bars, essentially making them tied for first among Chinese models. That judgment is closer to the chart than saying 'first'.

Tencent Hunyuan's official release mentions that Hy4 adds visual capabilities, covering scenarios such as finance, front-end, games, office work, and research, and also mentions considerations around cyber/biosecurity. The research page is here: https://hy.tencent.ai/research/hy4?langVersion=en

Some commenters mentioned that in Tencent's own financial evaluation, Hy4 scored 3.04 versus Kimi K3.1 Pro's 2.68, but the user-side results on Arena are almost even. AI Quanting replied that they would rather trust users, and that being even at this price is already good. It is not unusual for vendor evaluations and user-side evaluations to diverge, since task distributions differ.

There are two more caveats to remember. $0.57 refers to the median cost per task; model API unit pricing is a separate matter. @lin_builds asked in a reply how much of these 12K sessions was coding versus browser work, but the public information does not provide that distribution. For people preparing for production, these two categories affect decisions more than the ranking.

One reminder also appeared in the replies: Kimi K3.1 Pro will keep the open-model crown until Tencent puts the safetensors on Hugging Face. That statement clarifies the timing of 'first among open models'. Before the weights land, Hy4's first among Chinese models and 14th overall are established facts; first among open models has to wait for the next thing to happen. Leaderboard link: https://arena.ai/leaderboard/agent

To sum up, there are three things worth watching about Hy4 this time: a net improvement of +5.4% across 12K+ real Agent sessions, which did not drop as the sample grew; a median cost of $0.57 per task, with the gap versus Kimi K3.1 Pro within the margin of error; and first among Chinese models and 14th overall, the only Chinese lab in the top five. The caveats that need to be preserved are also clear: the weights have not yet been released, so first among open models must wait for the weights; the fight for first place looks more like a tie within the error bars. Next, watch the weight release, the task distribution, and independent reproduction.