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: 2026-10-16 13:02:49 UTC / 2026-10-16 21:02:49 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 | 1-6 | Claude Opus 5.5 (Max) 1 | 17.18%±3.38% | 19.36%±5.13% | 31.74%±13.06% | 17.94%±2.90% | 0.55%±0.11% | 6,796 |
| 2 | 1-7 | GPT 6.1 Sol (Max) 1 | 14.78%±2.03% | 15.69%±3.29% | 34.49%±7.44% | 14.73%±1.95% | 0.65%±0.04% | 12,705 |
| 3 | 1-7 | Claude Fable 5.1 (Max) 1 | 14.12%±2.01% | 15.71%±3.36% | 27.14%±7.30% | 14.70%±2.39% | 0.57%±0.11% | 16,282 |
| 4 | 1-7 | Gemini 4 Argon (High) 1 | 13.73%±1.52% | 14.58%±2.55% | 31.29%±5.30% | 16.78%±1.55% | 0.02%±0.55% | 28,097 |
| 5 | 1-7 | GPT 6 Astra (Max) 1 | 13.69%±2.32% | 12.55%±4.10% | 28.55%±8.52% | 14.56%±2.43% | 0.59%±0.09% | 12,892 |
| 6 | 1-9 | Claude Sonnet 5.5 (Max) 1 | 12.55%±3.13% | 19.98%±5.24% | 17.02%±11.59% | 12.35%±3.37% | 0.46%±0.24% | 7,446 |
| 7 | 2-9 | GPT 6 Sol (Max) 1 | 11.20%±2.27% | 9.21%±4.07% | 22.89%±8.17% | 12.41%±2.24% | 0.39%±0.10% | 7,075 |
| 8 | 6-12 | Claude Opus 5 (Max) 1 | 9.20%±1.69% | 12.65%±3.02% | 14.95%±5.75% | 7.26%±2.28% | 0.60%±0.08% | 23,614 |
| 9 | 6-12 | Claude Fable 5 (High) 1 | 8.39%±1.41% | 5.05%±2.68% | 15.07%±4.41% | 10.59%±2.10% | 0.62%±0.07% | 42,474 |
| 10 | 8-13 | GPT 5.6 Sol (xHigh) 1 | 7.42%±1.41% | 3.75%±2.70% | 14.75%±4.51% | 8.42%±2.08% | 0.51%±0.09% | 36,373 |
| 11 | 8-13 | Claude Opus 4.8 (High) 1 | 7.14%±1.53% | 3.96%±2.78% | 12.93%±4.77% | 7.02%±2.73% | 0.35%±0.19% | 40,277 |
| 12 | 8-18 | Muse Wonder (Max) 1 | 6.42%±2.15% | 8.43%±4.13% | 7.81%±7.39% | 8.68%±2.66% | 0.27%±0.30% | 5,526 |
| 13 | 10-23 | Claude Sonnet 5 (High) 1 | 4.69%±1.70% | 1.58%±3.43% | 4.74%±5.37% | 6.58%±2.46% | 0.41%±0.13% | 33,161 |
| 14 | 12-23 | Gemini 3.8 Flash (High) 1 | 3.97%±1.08% | 5.08%±2.23% | 4.55%±3.34% | 2.09%±1.60% | 0.52%±0.08% | 37,462 |
| 15 | 12-27 | MiMo V2.6 Pro 1 | 3.81%±1.18% | 8.61%±2.33% | 4.54%±3.85% | 2.00%±1.64% | 0.13%±0.27% | 23,636 |
| 16 | 12-25 | Muse Spark 1.3 (Max) 1 | 3.63%±0.81% | 4.87%±1.71% | 4.10%±2.57% | 5.60%±1.10% | 0.55%±0.06% | 73,452 |
| 17 | 12-31 | Grok 4.7 (xHigh) 1 | 3.39%±1.67% | 6.12%±3.42% | 3.48%±5.30% | 0.62%±2.55% | 0.56%±0.05% | 14,843 |
| 18 | 12-30 | GPT 5.5 (xHigh) 1 | 3.31%±1.28% | 1.98%±2.32% | 6.52%±3.62% | 3.20%±2.62% | 0.50%±0.14% | 54,310 |
| 19 | 13-25 | GPT 6 Luna (Max) 1 | 3.26%±0.49% | 1.30%±1.01% | 1.34%±1.57% | 6.17%±0.62% | 0.42%±0.06% | 155,783 |
| 20 | 13-27 | Deepseek V4.1 Flash (Max) 1 | 2.95%±0.37% | 7.48%±0.77% | 3.58%±1.19% | 1.94%±0.53% | 0.46%±0.05% | 228,658 |
| 21 | 13-29 | Hy4 preview 1 | 2.90%±0.69% | 6.04%±1.40% | 5.44%±2.23% | 2.43%±0.89% | 0.16%±0.17% | 82,690 |
| 22 | 13-33 | MiMo V2.6 Flash 1 | 2.53%±1.51% | 7.05%±2.98% | 3.55%±4.87% | 4.80%±1.97% | 3.18%±0.50% | 18,936 |
| 23 | 13-31 | GLM 5.2 (Max) 1 | 2.35%±0.83% | 3.55%±1.69% | 6.76%±2.50% | 1.23%±1.41% | 0.62%±0.05% | 86,843 |
| 24 | 15-33 | GLM 5.3 (Max) 1 | 2.00%±0.84% | 3.80%±1.75% | 6.09%±2.56% | 0.41%±1.31% | 0.59%±0.07% | 86,255 |
| 25 | 17-31 | Kimi K3 (Max) 1 | 1.95%±0.72% | 6.32%±1.50% | 2.21%±2.10% | 0.53%±1.25% | 0.65%±0.04% | 134,561 |
| 26 | 19-31 | Qwen3.8 Max 1 | 1.92%±0.61% | 4.69%±1.29% | 2.09%±1.87% | 1.91%±0.90% | 0.77%±0.21% | 67,520 |
| 27 | 15-33 | Claude Haiku 5.5 (High) 1 | 1.88%±0.98% | 1.64%±2.25% | 2.75%±3.07% | 6.40%±1.21% | 0.70%±0.47% | 21,839 |
| 28 | 17-33 | GPT 5.4 (High) 1 | 1.64%±1.09% | 4.28%±2.14% | 0.91%±3.15% | 7.10%±2.15% | 0.34%±0.19% | 81,779 |
| 29 | 19-35 | Grok 4.6 (xHigh) 1 | 1.16%±1.10% | 0.65%±2.39% | 0.32%±3.22% | 1.05%±1.76% | 0.52%±0.08% | 28,687 |
| 30 | 21-36 | GPT 5.6 Terra (xHigh) 1 | 0.43%±1.34% | 4.13%±2.73% | 1.96%±3.90% | 2.57%±2.26% | 0.40%±0.16% | 24,138 |
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。