Agent 智能體榜單

本頁顯示 Arena AI 前 30 名榜單嘅每日快照，方便快速了解近期模型表現。完整榜單、篩選項同最新變化請去 Arena AI 官網睇。

呢個榜單評測模型喺多輪工具調用／智能體任務上嘅綜合表現，涵蓋淨改善率、確認成功率、讚踩比、可控性、Bash 恢復率、工具幻覺率等維度。

數據更新時間: 2026-09-01 13:25:31 UTC / 2026-09-01 21:25:31 CST (北京时间)

排行榜反映特定評測同用戶投票偏好，唔等同於模型喺你嘅任務入面一定更好。揀模型時仲要考慮價格、速度、上下文、工具調用、私隱同地區可用性。

前 30 名

排名	排名區間	模型	淨改善率	確認成功率	讚踩比	可控性	Bash 恢復率	工具幻覺率	會話數
1	1-3	Claude Opus 5 (High) 1	13.77%±1.81%	15.59%±3.69%	21.87%±6.82%	15.56%±3.33%	15.00%±0.85%	0.82%±0.15%	21,433
2	1-6	Claude Opus 5 (Max) 1	11.61%±2.03%	15.64%±4.06%	18.91%±7.40%	7.19%±4.00%	15.47%±0.87%	0.86%±0.15%	17,073
3	1-7	Claude Fable 5 (High) 1	10.58%±1.54%	7.74%±3.28%	18.90%±5.48%	13.37%±2.85%	11.98%±2.07%	0.88%±0.14%	34,905
4	2-8	GPT 5.6 Sol (xHigh) 1	9.76%±1.53%	7.89%±3.19%	22.82%±5.69%	8.38%±2.88%	8.82%±1.03%	0.89%±0.14%	28,524
5	2-9	Claude Opus 4.8 (High) 1	9.51%±1.56%	5.86%±3.09%	20.70%±5.40%	11.46%±2.88%	9.83%±1.99%	0.28%±1.04%	36,887
6	3-8	Kimi K3 (Max) 1	8.74%±0.66%	16.89%±1.34%	16.20%±2.30%	1.85%±1.29%	7.86%±0.53%	0.89%±0.14%	94,549
7	4-16	GPT 5.5 (xHigh) 1	7.89%±1.07%	2.61%±2.34%	13.34%±3.78%	8.74%±1.98%	13.88%±1.19%	0.88%±0.14%	50,040
8	2-19	Claude Sonnet 5 (High) 1	7.48%±2.17%	0.98%±4.54%	12.24%±7.47%	12.40%±4.56%	11.05%±1.64%	0.74%±0.17%	27,467
9	6-19	Claude Opus 4.7 (High) 1	6.64%±1.40%	4.09%±2.94%	10.61%±4.65%	6.39%±2.60%	11.29%±2.35%	0.80%±0.16%	36,570
10	7-20	Claude Opus 4.7 1	6.33%±1.42%	3.65%±3.05%	10.63%±4.65%	7.86%±2.61%	8.66%±2.49%	0.83%±0.15%	37,121
11	7-19	GPT 5.5 (High) 1	6.10%±1.00%	1.68%±2.13%	10.65%±3.46%	5.87%±1.80%	11.40%±1.53%	0.89%±0.14%	73,880
12	7-19	GLM 5.2 (Max) 1	6.08%±0.78%	8.49%±1.68%	10.59%±2.71%	5.45%±1.43%	4.99%±0.81%	0.89%±0.14%	65,079
13	7-20	Grok 4.5 1	6.06%±1.15%	6.01%±2.55%	5.63%±4.09%	7.20%±2.24%	10.57%±1.17%	0.89%±0.14%	34,008
14	7-20	Qwen3.8 Max 1	6.01%±1.13%	10.76%±2.43%	6.80%±3.91%	5.40%±2.25%	7.18%±1.00%	0.11%±0.27%	18,386
15	7-20	DeepSeek V4 Pro (High) (0813) 1	5.90%±1.09%	12.90%±2.25%	5.39%±3.72%	1.23%±2.43%	9.10%±0.89%	0.89%±0.14%	23,928
16	7-21	Grok 4.6 (xHigh) 1	5.76%±1.29%	11.89%±2.95%	2.52%±4.44%	3.80%±2.74%	9.70%±0.90%	0.89%±0.14%	15,090
17	8-24	Claude Opus 4.6 1	5.22%±1.34%	3.13%±2.96%	7.52%±4.49%	5.24%±2.48%	9.31%±1.84%	0.88%±0.14%	36,365
18	8-24	GPT 5.5 1	4.84%±0.88%	2.11%±1.95%	6.32%±2.97%	4.76%±1.61%	10.13%±1.17%	0.89%±0.14%	77,762
19	8-27	GLM 5.3 Flash 1	4.41%±1.37%	15.21%±2.85%	5.21%±4.56%	1.89%±2.96%	1.15%±1.28%	0.89%±0.14%	9,970
20	16-27	GLM 5.3 (Max) 1	3.82%±0.80%	12.60%±1.78%	4.23%±2.62%	0.73%±1.60%	0.65%±0.85%	0.89%±0.14%	41,022
21	17-29	GPT 5.4 (High) 1	3.25%±0.86%	2.52%±2.04%	0.11%±2.83%	4.37%±1.72%	8.57%±0.94%	0.89%±0.14%	76,973
22	19-30	Deepseek V4 Flash (High) (20260731) 1	2.97%±0.79%	7.45%±1.81%	2.61%±2.58%	0.70%±1.62%	3.24%±0.67%	0.87%±0.15%	51,219
23	17-32	GPT 5.6 Terra (xHigh) 1	2.92%±1.18%	3.10%±3.03%	1.02%±3.84%	9.42%±2.37%	8.40%±1.29%	0.89%±0.14%	16,778
24	17-33	Qwen3.8 Flash Next 1	2.44%±1.88%	12.34%±3.72%	1.59%±6.40%	1.06%±4.39%	2.92%±1.03%	0.41%±0.35%	8,777
25	12-34	Claude Opus 4.8 1	2.33%±2.60%	6.69%±3.19%	11.68%±5.32%	9.92%±2.97%	10.51%±1.84%	27.15%±10.47%	33,690
26	19-33	GPT 5.6 Luna (xHigh) 1	2.13%±1.23%	0.74%±2.92%	0.13%±3.98%	2.42%±2.57%	8.22%±1.21%	0.89%±0.14%	16,383
27	21-33	Qwen 3.8 27B 1	1.50%±1.17%	7.24%±2.67%	1.39%±3.90%	0.20%±2.39%	1.49%±1.05%	0.16%±0.25%	16,060
28	19-37	Kimi K2.7 Code 1	1.38%±2.48%	2.15%±5.45%	1.19%±8.44%	5.26%±5.45%	2.56%±2.76%	0.89%±0.14%	11,142
29	23-34	Muse Spark 1.2 (xHigh) 1	0.97%±1.12%	6.38%±2.81%	6.40%±3.36%	5.49%±2.42%	9.47%±1.38%	0.88%±0.15%	18,416
30	23-33	Gemini 3.7 Flash (High) 1	0.96%±0.90%	9.83%±2.11%	1.60%±2.91%	5.19%±1.89%	0.93%±1.05%	0.82%±0.15%	24,413

點樣睇呢張表

• 排名 / 排名區間：Arena AI 根據智能體任務對戰投票估算嘅相對名次；排名區間表示喺置信範圍內嘅名次波動。
• 淨改善率：相比基線喺多輪工具調用／智能體任務中嘅淨改善幅度。
• 確認成功率：任務完成後用戶確認成功（讚）嘅比例。
• 讚踩比：用戶對回答投「讚」同「踩」嘅比例，反映主觀滿意度。
• 可控性：模型遵循指令並保持穩定輸出嘅能力。
• Bash 恢復率：喺 Bash／命令列操作出錯後能自我恢復嘅比例。
• 工具幻覺率：模型虛構或調用不存在工具嘅比例，越低越好。
• 會話數：樣本量參考；樣本較少時，排名通常更容易波動。

揀模型時再確認三件事

1. Provider 係咪真係有提供呢個模型，以及你嘅地區同帳戶係咪可用；
2. API 價格、速率限制同上下文係咪適合你嘅任務；
3. 用 3～5 個真實任務做小規模測試，唔好只睇總榜名次。

數據來源

數據嚟自 Arena AI 官方 Agent 智能體榜單，由 GitHub Actions 每日更新。模型價格、授權同能力請以模型供應商官方資訊為準。