Arena says Qwen4-Max posted a +7.5% net uplift across 14.7k real Agent sessions, entering the Agent Arena overall leaderboard at No. 14 for the first time. The post also lists Qwen3.8-Max at No. 39. The rank change and net uplift are two results reported by the post; they do not establish a direct paired win rate or an increase in task success rate.
Suitable for: Understanding the Agent Arena overall leaderboard position and session count published by Arena.
Not suitable for extrapolating to: Inferring success rates for a specific workflow, cost-effectiveness, or direct comparisons with other testing frameworks.
Applicable model version: Qwen4-Max; the post lists Qwen3.8-Max as a reference ranking at No. 39.
Test environment or client: Arena Agent Arena; the post says the data comes from real Agent session tracking.
Reasoning tier and parameters: Not specified.
The post says the results are based on 14.7k real Agent sessions and real task traces from the global community. It does not specify how sessions were filtered, the test window, Agent configuration, statistical method, or uncertainty intervals. It provides no task records sufficient to reproduce the calculation.
Qwen4-Max: No. 14 on the Agent Arena overall leaderboard, with a +7.5% net uplift.
The post lists the previous model, Qwen3.8-Max, at No. 39 and Qwen4-Max at No. 14. It gives no head-to-head win rate.
The post also reports positive gains across four signals and three categories, but does not provide the underlying tables, denominators, or interval estimates here.
| Item reported in the post | Result |
|---|---|
| Number of real Agent sessions | 14.7k |
| Qwen4-Max overall ranking | #14 |
| Qwen4-Max net uplift | +7.5% |
| Qwen3.8-Max overall ranking | #39 |
The original leaderboard is at https://arena.ai/leaderboard/agent. At the time of collection, the page returned a Cloudflare 403, so the dynamic leaderboard snapshot, costs, and other metrics in its tables could not be checked. This note records only the visible content of the X post.
Arena's post gives the leaderboard rank, net uplift, and session count at the time of publication. The underlying leaderboard and methodology are needed to verify rank uncertainty, how net uplift is calculated, and the cost per task.
Open the original post in Tabbit and read its public content. The original leaderboard page returned 403 at the time of collection, and access restrictions were not bypassed; therefore, the post's net uplift could not be independently recalculated, nor could the dynamic ranking be verified.
Qwen4-Max