Arena Agent · 工具幻觉
调用根本不存在的工具的情况,分越高越少乱编 完整榜单 53 个模型,样本量 211,147,826 次观测。
综合总榜535 个行为信号加权后的总排名任务确认完成53用户在会话里明确确认「任务完成了」的比例好评 vs 差评53用户明确表扬多于明确抱怨的程度可操控性53用户纠正、反驳之后,模型能改到点子上的比例Bash Reliability53How often the model’s bash commands succeed. Errors outside its control aren’t counted. E.g. network issues, permission denied, etc.工具幻觉53调用根本不存在的工具的情况,分越高越少乱编
| Rank | Model | 信号分 |
|---|---|---|
1 | GPT 6.1 Sol (Max) | +0.6 |
2 | Kimi K3 (Max) | +0.6 |
3 | Step 5 Preview | +0.6 |
4 | GLM 5.2 (Max) | +0.6 |
5 | Gemini 3.6 Flash (High) | +0.6 |
6 | Claude Fable 5 (High) | +0.6 |
7 | Claude Opus 5 (Max) | +0.6 |
8 | GLM 5.3 (Max) | +0.6 |
9 | GPT 6 Astra (Max) | +0.6 |
10 | Grok 4.5 | +0.6 |
11 | GLM 5.3 Flash | +0.6 |
12 | Claude Fable 5.1 (Max) | +0.6 |
13 | Grok 4.7 (xHigh) | +0.6 |
14 | Muse Spark 1.3 (Max) | +0.5 |
15 | Claude Opus 5.5 (Max) | +0.5 |
16 | Grok 4.6 (xHigh) | +0.5 |
17 | Gemini 3.8 Flash (High) | +0.5 |
18 | GPT 5.6 Sol (xHigh) | +0.5 |
19 | GPT 5.5 (xHigh) | +0.5 |
20 | DeepSeek V4 Pro (High) (0813) | +0.5 |
21 | Deepseek V4.1 Flash (Max) | +0.5 |
22 | Claude Sonnet 5.5 (Max) | +0.5 |
23 | GPT 5.5 | +0.4 |
24 | GPT 5.6 Luna (xHigh) | +0.4 |
25 | GPT 6 Luna (Max) | +0.4 |
26 | Claude Sonnet 5 (High) | +0.4 |
27 | Gemini 3.7 Flash (High) | +0.4 |
28 | GPT 5.6 Terra (xHigh) | +0.4 |
29 | GPT 6 Sol (Max) | +0.4 |
30 | Claude Opus 4.8 (High) | +0.3 |
31 | GPT 5.4 (High) | +0.3 |
32 | Muse Spark 1.2 (xHigh) | +0.2 |
33 | Hy4 preview | +0.1 |
34 | Muse Spark 1.1 | +0.1 |
35 | Qwen3.7 Max | +0.1 |
36 | Gemini 4 Argon (High) | -0.0 |
37 | Minimax M3 | -0.1 |
38 | MiMo V2.6 Pro | -0.1 |
39 | Qwen 3.8 27B | -0.2 |
40 | Gemini 3.1 Pro Preview | -0.2 |
41 | Inkling | -0.5 |
42 | Qwen3.8 Flash Next | -0.7 |
43 | Claude Haiku 5.5 (High) | -0.7 |
44 | Qwen3.8 Max | -0.8 |
45 | Nemotron 3 Ultra | -0.9 |
46 | Mimo V2.5 Pro | -1.1 |
47 | Inkling Small | -1.2 |
48 | Qwen3.7 Plus | -1.2 |
49 | Solar Pro 4 | -2.1 |
50 | MiMo V2.6 Flash | -3.2 |
51 | Mistral Medium 3.5 | -3.2 |
52 | Hy3 | -4.2 |
53 | Mistral Large 4 | -17.5 |
数据更新: 2026-10-14数据来源: Arena AI ↗