Agent Arena 智能体能力排行榜
Arena 的 Agent Arena 让真实用户在 Agent Mode 中与随机分配的模型长时间协作,再用因果追踪估计每个模型作为智能体编排者相对平均模型的净提升。分数可正可负,正值代表优于平均水平。
榜首模型
Claude Opus 5.5 (Max) · +23.35
会话数
851,360
模型数
55
数据版本
2026-10-23
数据来源: Arena Agent Arena
这个榜单怎么读
真实会话而非两两对战:每个 Agent Mode 会话随机分配一个编排模型,用户与单个智能体持续交互,有时长达数百轮,不是看两个匿名回答投票。
因果追踪而非 Elo:对每个行为信号估计“换成这个模型”相对平均模型的处理效应(百分点)。综合分是各信号的加权平均,其中 Bash 可靠性按半权重计。
相对且动态:基线是当前所有模型的平均水平,新模型加入会让所有分数随之变化;置信区间重叠较多的模型应视为接近,名次区间给出统计上可能的名次范围。
评分信号
- 确认成功
- 用户明确确认任务已完成的频率。
- 赞扬 vs 抱怨
- 任务过程中用户的正面表达是否多于负面表达。
- 可引导性
- 用户纠正或反驳后,模型的下一次回复能否被接受。
- Bash 可靠性(半权重)
- Shell 命令执行成功的可靠性,网络、权限等非模型原因的失败不计入;综合分中按半权重计。
- 工具幻觉控制
- 调用不存在工具的情况;数值已统一为越高越好,即幻觉越少。
模型排名
| 排名 | 模型 | 净提升 | 确认成功 | 赞扬 vs 抱怨 | 可引导性 | Bash 可靠性 | 工具幻觉控制 | 会话数 | 单任务成本 P50 | 单任务耗时 P50 |
|---|---|---|---|---|---|---|---|---|---|---|
| 11–3 | Claude Opus 5.5 (Max)Anthropic | +23.35±5.32 | +31.07±5.72 | +51.88±21.52 | +13.99±3.65 | +15.45±1.07 | +0.40±0.19 | 3,287 | $3.51 | 9.8 分钟 |
| 21–9 | Claude Sonnet 5.5 (Max)Anthropic | +17.07±6.34 | +24.71±8.63 | +37.26±24.68 | +7.40±6.19 | +13.91±2.05 | +0.50±0.06 | 2,487 | $2.28 | 11.4 分钟 |
| 31–9 | Claude Fable 5.1 (Max)Anthropic | +15.23±3.57 | +18.22±5.32 | +32.00±13.58 | +10.90±3.17 | +13.81±2.17 | +0.49±0.06 | 6,246 | $3.81 | 5.1 分钟 |
| 42–9 | Gemini 4 Argon (High)Google | +14.79±1.96 | +19.74±3.15 | +30.22±7.43 | +14.31±1.72 | +4.16±2.08 | +0.18±0.23 | 11,324 | $0.539 | 3.5 分钟 |
| 52–14 | GPT-6 Astra (Max)OpenAI | +11.71±3.88 | +18.80±6.26 | +14.30±14.14 | +11.86±3.68 | +14.46±2.63 | +0.50±0.06 | 4,779 | $2.00 | 5.1 分钟 |
| 62–13 | GPT-6.1 Sol (Max)OpenAI | +11.65±3.38 | +12.49±5.72 | +25.24±12.69 | +9.31±2.91 | +9.74±2.28 | +0.50±0.06 | 4,442 | $0.441 | 4.6 分钟 |
| 72–14 | Muse Wonder Full (Max)Meta | +10.83±2.90 | +14.06±5.04 | +21.25±10.61 | +8.25±3.06 | +9.75±1.54 | +0.31±0.26 | 4,185 | $0.365 | 5.8 分钟 |
| 82–14 | Claude Opus 5 (Max)Anthropic | +9.55±3.39 | +12.84±5.37 | +18.19±12.32 | +4.62±3.59 | +13.69±1.79 | +0.50±0.06 | 9,138 | $2.30 | 5.1 分钟 |
| 95–14 | Gemini 3.8 Flash (High)Google | +8.17±2.24 | +11.00±3.75 | +13.87±8.14 | +5.39±2.35 | +12.53±1.39 | +0.24±0.31 | 11,647 | $0.411 | 4.3 分钟 |
| 102–30 | Kimi K3.1 Pro (Max)初步结果Moonshot | +7.69±5.95 | +1.09±11.64 | +23.94±22.54 | +4.46±5.78 | +9.20±3.36 | +0.50±0.06 | 1,152 | $0.618 | 3.8 分钟 |
| 115–19 | Claude Fable 5 (High)Anthropic | +6.66±2.60 | +5.84±4.54 | +12.89±8.66 | +3.24±3.49 | +15.00±2.56 | +0.50±0.06 | 17,019 | $1.84 | 2.2 分钟 |
| 125–21 | GPT-5.6 Sol (xHigh)OpenAI | +6.13±2.77 | +2.39±5.05 | +14.99±9.43 | +5.17±3.17 | +9.69±2.02 | +0.18±0.33 | 14,354 | $0.828 | 2.5 分钟 |
| 135–25 | GPT-6 Sol (Max)OpenAI | +5.88±3.31 | +8.31±6.40 | +5.47±11.53 | +5.46±3.51 | +13.74±1.39 | +0.38±0.11 | 3,169 | $0.933 | 3.6 分钟 |
| 146–31 | +4.79±3.45 | +12.13±5.97 | +3.90±12.23 | +0.58±3.69 | +9.08±1.51 | +0.42±0.09 | 5,285 | $1.21 | 6.6 分钟 | |
| 1510–27 | MiMo-V2.6-ProXiaomi | +4.01±1.70 | +8.48±3.17 | +5.30±5.84 | +1.68±1.91 | +6.00±1.31 | -0.41±0.59 | 12,630 | $0.094 | 7.6 分钟 |
| 1610–32 | Gemini 3.7 Flash (High)Google | +3.06±1.84 | +3.59±3.33 | +3.95±6.00 | +1.64±2.43 | +8.56±3.27 | +0.34±0.32 | 20,987 | $0.264 | 2.8 分钟 |
| 1710–33 | Muse Spark 1.3 (Max)Meta | +2.58±1.68 | +3.08±3.19 | +4.56±5.75 | +1.80±1.93 | +3.40±1.82 | +0.48±0.06 | 26,928 | $0.309 | 2.7 分钟 |
| 1810–34 | DeepSeek-V4.1-Flash-20261017 (Max)DeepSeek | +2.47±1.82 | +7.12±3.61 | +0.70±6.26 | +0.83±2.02 | +4.22±1.06 | +0.35±0.12 | 9,680 | $0.115 | 3.6 分钟 |
| 1911–33 | Hy4 previewTencent | +2.17±1.36 | +5.30±2.60 | +5.70±4.76 | +0.25±1.46 | -3.24±1.07 | +0.14±0.21 | 27,293 | $0.293 | 8.8 分钟 |
| 2010–38 | Muse Wonder (Max)Meta | +1.95±3.23 | +5.94±6.66 | -3.78±10.71 | +2.07±3.92 | +8.10±2.45 | +0.47±0.09 | 3,307 | $0.933 | 2.5 分钟 |
| 2113–33 | DeepSeek-V4.1-Flash-20260910 (Max)DeepSeek | +1.63±0.73 | +4.91±1.38 | +3.05±2.51 | -0.46±0.87 | -1.03±0.66 | +0.34±0.07 | 76,799 | $0.100 | 3.3 分钟 |
| 2212–35 | Kimi K3 (Max)Moonshot | +1.43±1.55 | +4.45±2.90 | +3.38±4.99 | -0.15±2.10 | -3.45±3.02 | +0.50±0.06 | 52,607 | $0.644 | 5.8 分钟 |
| 2314–35 | GPT-6 Luna (Max)OpenAI | +1.18±0.89 | -0.30±1.78 | -2.60±3.01 | +2.84±0.98 | +9.96±0.73 | +0.38±0.09 | 52,618 | $0.070 | 3.6 分钟 |
| 2414–38 | GLM-5.3 (Max)Z.ai | +0.39±1.64 | -0.54±3.19 | +5.56±5.36 | -1.99±1.97 | -3.59±2.64 | +0.50±0.06 | 32,296 | $0.461 | 6.2 分钟 |
| 2512–40 | GPT-5.6 Terra (xHigh)OpenAI | +0.33±2.63 | -3.53±4.73 | +2.40±9.09 | +0.61±3.38 | +3.00±4.27 | +0.50±0.06 | 9,476 | $0.427 | 1.6 分钟 |
| 2615–36 | Qwen3.8-MaxAlibaba | +0.32±1.22 | +4.25±2.40 | -1.37±3.99 | +0.99±1.49 | -3.01±1.80 | -0.91±0.52 | 25,434 | $0.634 | 6.7 分钟 |
| 2713–39 | +0.28±2.25 | -3.16±4.29 | -1.13±7.30 | +1.62±2.56 | +7.19±2.70 | +0.32±0.18 | 11,230 | $0.841 | 4.4 分钟 | |
| 2812–40 | MiMo-V2.6-FlashXiaomi | +0.20±2.78 | +4.17±5.19 | -3.33±9.24 | +2.92±2.99 | -1.69±2.64 | -2.03±1.18 | 7,009 | $0.049 | 5.9 分钟 |
| 2912–42 | Claude Sonnet 5 (High)Anthropic | +0.02±3.23 | -4.91±6.51 | -4.49±10.50 | +2.08±4.02 | +14.00±2.54 | +0.43±0.14 | 13,078 | $0.573 | 1.8 分钟 |
| 3017–39 | GLM-5.3-FlashZ.ai | -0.30±1.40 | +3.15±2.81 | -0.71±4.55 | -0.27±1.70 | -7.83±2.28 | +0.37±0.10 | 31,428 | $0.082 | 3.1 分钟 |
| 3120–40 | GLM-5.2 (Max)Z.ai | -0.96±1.66 | -3.64±3.20 | +3.57±5.33 | -2.29±2.27 | -4.68±2.73 | +0.39±0.13 | 34,164 | $0.494 | 5.5 分钟 |
| 3211–50 | Beam 1.0 501B A23BReflection | -1.48±5.06 | -0.43±8.27 | -1.81±18.20 | -6.55±5.08 | +4.04±5.04 | +0.13±0.38 | 2,905 | $0.519 | 34 秒 |
| 3314–48 | DeepSeek-V4-Pro (High, 20260813)DeepSeek | -1.70±3.53 | -8.94±7.38 | +0.56±11.57 | +1.52±3.79 | -1.84±3.22 | +0.14±0.50 | 2,337 | — | — |
| 3421–47 | Gemini 3.1 Pro PreviewGoogle | -2.02±2.40 | -13.46±4.49 | -0.58±7.21 | +0.05±3.86 | +8.93±4.73 | +0.45±0.09 | 30,000 | $0.369 | 2.0 分钟 |
| 3526–45 | Muse Spark 1.1Meta | -2.60±1.22 | +0.60±2.41 | -9.99±3.58 | -5.15±1.98 | +5.51±2.70 | +0.06±0.16 | 56,059 | $0.242 | 56 秒 |
| 3616–50 | MiniMax-M3.1-Flash初步结果MiniMax | -2.92±4.24 | +0.32±8.48 | -1.43±13.92 | -5.71±5.61 | -13.35±3.59 | +0.34±0.17 | 2,734 | $0.129 | 3.8 分钟 |
| 3724–47 | GPT-6 Luna (Medium)OpenAI | -2.97±1.89 | -6.40±4.05 | -6.33±5.97 | -2.19±2.30 | +2.89±4.01 | +0.09±0.51 | 8,349 | $0.020 | 59 秒 |
| 3824–49 | Claude Haiku 5.5 (High)Anthropic | -3.34±2.12 | -2.31±4.94 | -14.27±6.53 | -2.17±2.58 | +7.78±1.87 | -0.16±0.51 | 7,887 | $0.045 | 58 秒 |
| 3923–50 | Mistral Large 4Mistral | -3.88±3.23 | +0.79±7.12 | -5.14±9.88 | +0.98±3.98 | +3.69±3.66 | -15.96±3.43 | 4,140 | $0.406 | 2.2 分钟 |
| 4031–49 | Muse Spark 1.2 (xHigh)Meta | -4.69±1.67 | -4.83±3.57 | -10.68±5.09 | -3.39±2.11 | -4.26±3.58 | -0.06±0.26 | 19,214 | $0.242 | 57 秒 |
| 4132–49 | Qwen3.8-Flash-NextAlibaba | -4.91±1.43 | -0.63±2.86 | -4.63±4.62 | -4.80±1.72 | -23.14±1.91 | -0.46±0.24 | 33,307 | $0.144 | 4.7 分钟 |
| 4228–50 | Step 5 PreviewStepFun | -5.03±2.86 | -5.18±6.47 | -12.51±9.00 | -0.89±3.59 | -8.51±2.94 | +0.23±0.44 | 4,755 | $0.269 | 4.2 分钟 |
| 4332–50 | GPT-6 Luna (Low)OpenAI | -5.53±2.16 | -11.19±4.67 | -9.55±6.66 | -4.41±2.58 | -0.45±4.38 | +0.50±0.06 | 5,462 | $0.009 | 30 秒 |
| 4431–51 | Qwen3.7-PlusAlibaba | -5.82±2.95 | -6.55±6.01 | -12.53±9.04 | -3.89±4.33 | -4.12±6.63 | -1.14±1.55 | 9,071 | $0.089 | 2.1 分钟 |
| 4533–50 | Qwen3.8-27BAlibaba | -5.87±1.56 | -3.43±3.16 | -9.26±4.87 | -4.70±2.08 | -17.48±3.67 | -0.28±0.35 | 21,634 | $0.244 | 4.2 分钟 |
| 4632–50 | Qwen3.7-MaxAlibaba | -6.08±2.32 | -15.34±4.80 | -8.83±6.91 | -3.94±3.26 | +1.86±6.70 | -0.16±0.54 | 16,255 | $0.194 | 1.0 分钟 |
| 4733–51 | MiMo-V2.5-ProXiaomi | -6.55±2.18 | -12.94±4.36 | -10.44±6.36 | -6.67±3.09 | +2.64±5.96 | -0.73±0.72 | 17,709 | $0.058 | 2.5 分钟 |
| 4835–54 | Mistral Medium 3.5Mistral | -8.70±3.52 | -18.39±7.02 | -13.57±10.35 | -3.40±4.44 | -1.78±9.12 | -2.89±2.38 | 4,710 | $0.226 | 1.3 分钟 |
| 4939–54 | Hy3Tencent | -9.03±2.58 | -15.35±5.14 | -12.15±7.28 | -4.15±3.61 | -16.51±8.40 | -0.73±0.57 | 11,670 | $0.050 | 1.8 分钟 |
| 5036–55 | NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4)Nvidia | -9.69±4.40 | -14.04±9.92 | -18.14±12.51 | -4.86±5.11 | -12.23±12.58 | -0.47±1.36 | 1,402 | $0.247 | 2.2 分钟 |
| 5146–54 | MiniMax-M3MiniMax | -10.74±2.18 | -19.61±4.86 | -18.06±5.88 | -8.37±3.39 | -4.05±5.60 | -0.27±0.75 | 16,716 | $0.084 | 51 秒 |
| 5248–55 | InklingThinky | -11.98±3.08 | -24.13±6.12 | -22.02±7.63 | -4.59±3.92 | -4.14±14.39 | -1.09±1.28 | 15,738 | $0.203 | 1.1 分钟 |
| 5348–55 | Inkling-SmallThinky | -14.91±3.30 | -28.64±8.01 | -24.90±7.93 | -15.43±5.15 | +4.64±7.64 | -0.46±0.84 | 4,965 | $0.094 | 1.5 分钟 |
| 5448–55 | Nemotron 3 UltraNvidia | -15.21±6.22 | -25.19±13.75 | -25.55±15.62 | -19.23±12.82 | +3.32±13.17 | -0.14±0.44 | 4,882 | $0.083 | 1.6 分钟 |
| 5551–55 | Solar Pro 4Upstage | -17.62±3.94 | -33.62±10.12 | -18.26±11.12 | -10.03±5.05 | -30.42±11.83 | -2.16±2.13 | 3,981 | $0.020 | 42 秒 |
分数为相对平均模型的百分点差值,所有信号均已统一为数值越高越好;± 为 95% 置信区间。成本与耗时为近 14 天单任务中位数。












