智能体榜单
本页展示 Arena AI 榜单前 30 名的最新快照，方便快速了解近期模型表现。完整榜单、筛选项和最新变化请到 Arena AI 官网查看。
本榜单评测模型在多轮工具调用 / 智能体任务上的综合表现，涵盖净改进率、确认成功率、赞踩比、可控性、Bash 恢复率、工具幻觉率等维度。
数据更新时间: 2026-11-04 16:05:06 UTC / 2026-11-05 00:05:06 CST (北京时间)
前 30 名
| 排名 | 排名区间 | 模型 | 净改进率 | 确认成功率 | 赞踩比 | 可控性 | 工具幻觉率 | 会话数 |
|---|---|---|---|---|---|---|---|---|
| 1 | 1-2 | Claude Opus 5.5 (Max) 1 | 20.02%±2.08% | 20.75%±3.23% | 42.34%±7.91% | 18.90%±1.86% | 0.59%±0.18% | 14,089 |
| 2 | 2-4 | Gemini 4 Carbon (High) 1 | 16.54%±0.78% | 18.04%±1.27% | 39.87%±2.85% | 18.98%±0.74% | 0.44%±0.21% | 42,599 |
| 3 | 1-6 | Claude Sonnet 5.5 (Max) 1 | 15.41%±2.81% | 19.95%±4.43% | 30.05%±10.89% | 12.33%±2.99% | 0.72%±0.12% | 8,516 |
| 4 | 2-6 | GPT-6 Astra (Max) 1 | 14.46%±2.04% | 14.28%±3.36% | 31.72%±7.50% | 14.14%±1.95% | 0.83%±0.06% | 15,391 |
| 5 | 3-7 | Claude Fable 5.1 (Max) 1 | 13.42%±1.89% | 12.75%±3.57% | 26.75%±6.95% | 15.20%±1.95% | 0.72%±0.12% | 17,496 |
| 6 | 3-8 | GPT-6.1 Sol (Max) 1 | 12.46%±1.10% | 12.22%±1.95% | 26.93%±3.97% | 13.50%±1.11% | 0.83%±0.06% | 30,088 |
| 7 | 5-9 | GPT-6.1 Terra (Max) 1 | 10.91%±1.38% | 9.34%±2.82% | 17.11%±4.81% | 15.00%±1.35% | 0.83%±0.06% | 20,731 |
| 8 | 6-9 | Gemini 4 Argon (High) 1 | 10.60%±0.85% | 12.82%±1.55% | 23.71%±2.91% | 13.61%±0.96% | 0.38%±0.13% | 52,536 |
| 9 | 7-11 | Muse Wonder Full (Max) 1 | 9.59%±0.98% | 11.60%±1.80% | 17.07%±3.09% | 10.73%±1.05% | 0.36%±1.59% | 31,819 |
| 10 | 9-12 | Claude Fable 5 (High) 1 | 7.90%±1.11% | 5.58%±2.18% | 15.40%±3.61% | 8.95%±1.38% | 0.55%±0.19% | 51,186 |
| 11 | 9-13 | Claude Opus 5 (Max) 1 | 7.71%±1.24% | 11.78%±2.34% | 11.90%±4.13% | 5.22%±1.60% | 0.78%±0.08% | 32,011 |
| 12 | 10-15 | GPT-6 Sol (Max) 1 | 6.62%±1.56% | 3.44%±3.45% | 9.75%±5.15% | 10.10%±1.75% | 0.83%±0.06% | 12,598 |
| 13 | 12-18 | Kimi K3.1 Pro (Max) 1 | 5.09%±1.30% | 3.94%±2.90% | 7.68%±4.09% | 6.50%±1.64% | 0.83%±0.06% | 17,568 |
| 14 | 11-22 | Hy4 1 | 4.66%±2.08% | 8.66%±3.96% | 3.77%±7.17% | 5.66%±3.03% | 0.21%±0.37% | 8,451 |
| 15 | 12-19 | GPT-5.6 Sol (xHigh) 1 | 4.59%±1.04% | 1.97%±2.24% | 5.79%±3.25% | 6.44%±1.39% | 0.83%±0.06% | 45,488 |
| 16 | 13-23 | MiMo-V2.6-Pro 1 | 2.93%±0.92% | 5.84%±1.92% | 5.27%±2.83% | 1.08%±1.25% | 0.15%±0.24% | 48,399 |
| 17 | 13-31 | Grok 4.7 (xHigh) 1 | 2.71%±1.48% | 4.73%±3.13% | 3.30%±4.69% | 1.38%±2.05% | 0.83%±0.06% | 18,860 |
| 18 | 15-24 | Gemini 3.8 Flash (High) 1 | 2.63%±0.82% | 1.34%±1.76% | 5.28%±2.50% | 0.12%±1.17% | 0.35%±0.15% | 53,021 |
| 19 | 14-31 | Claude Sonnet 5 (High) 1 | 2.58%±1.18% | 0.08%±2.59% | 0.40%±3.57% | 4.01%±1.72% | 0.53%±0.12% | 41,798 |
| 20 | 13-31 | Muse Wonder (Max) 1 | 2.52%±1.36% | 6.75%±2.92% | 0.14%±4.08% | 3.10%±1.84% | 0.82%±1.65% | 13,176 |
| 21 | 15-25 | Muse Spark 1.3 (Max) 1 | 2.49%±0.83% | 3.75%±1.80% | 2.11%±2.53% | 3.60%±1.11% | 0.78%±0.07% | 80,963 |
| 22 | 15-25 | DeepSeek-V4.1-Flash-20261017 (Max) 1 | 2.25%±0.57% | 3.60%±1.33% | 1.53%±1.76% | 3.04%±0.76% | 0.32%±0.09% | 68,771 |
| 23 | 16-29 | GPT-6 Luna (Max) 1 | 1.90%±0.43% | 0.19%±0.99% | 1.15%±1.34% | 4.19%±0.57% | 0.83%±0.06% | 225,233 |
| 24 | 17-32 | GLM-5.3 (Max) 1 | 1.12%±0.80% | 3.41%±1.73% | 4.29%±2.38% | 0.80%±1.13% | 0.83%±0.06% | 96,259 |
| 25 | 18-32 | Hy4 preview 1 | 1.04%±0.70% | 4.93%±1.54% | 1.02%±2.18% | 0.21%±0.98% | 0.07%±0.18% | 101,091 |
| 26 | 20-32 | Kimi K3 (Max) 1 | 0.91%±0.58% | 3.89%±1.29% | 1.30%±1.73% | 0.36%±0.84% | 0.83%±0.06% | 161,198 |
| 27 | 20-32 | GLM-5.2 (Max) 1 | 0.75%±0.77% | 2.60%±1.68% | 2.71%±2.32% | 0.54%±1.19% | 0.83%±0.06% | 99,220 |
| 28 | 20-32 | MiMo-V2.6-Flash 1 | 0.51%±1.04% | 5.42%±2.27% | 0.60%±3.20% | 2.51%±1.39% | 2.97%±0.39% | 30,432 |
| 29 | 24-32 | Qwen3.8-Max 1 | 0.50%±0.65% | 4.28%±1.42% | 0.27%±1.98% | 0.14%±0.95% | 0.74%±0.20% | 79,954 |
| 30 | 21-32 | Claude Haiku 5.5 (High) 1 | 0.39%±1.03% | 1.29%±2.50% | 5.12%±3.09% | 3.24%±1.27% | 1.33%±0.66% | 30,908 |
怎么看这张表
- 排名 / 排名区间：Arena AI 根据智能体任务对战投票估算的相对名次；排名区间表示在置信范围内的名次波动。
- 净改进率：相比基线在多轮工具调用 / 智能体任务中的净改进幅度。
- 确认成功率：任务完成后用户确认成功（赞）的比例。
- 赞踩比：用户对回答投“赞”与“踩”的比例，反映主观满意度。
- 可控性：模型遵循指令并保持稳定输出的能力。
- Bash 恢复率：在 Bash / 命令行操作出错后能自我恢复的比例。
- 工具幻觉率：模型虚构或调用不存在工具的比例，越低越好。
- 会话数：样本量参考；样本较少时，排名通常更容易波动。
选模型时再确认三件事
- Provider 是否实际提供这个模型，以及你的地区和账户是否可用；
- API 价格、速率限制和上下文是否适合你的任务；
- 用 3～5 个真实任务做小规模测试，不要只看总榜名次。
数据来源
数据来自 Arena AI 官方 智能体榜单，由 GitHub Actions 每天更新一次。模型价格、许可证和能力请以模型服务商官方信息为准。