Agent 对齐指数排行榜 - 2026年10月最新对齐指数与失败率

排名	排名区间排名区间：基于 Elo 分数估算出的模型在不同对决中可能变动的最乐观与最保守排名范围	模型	对齐指数对齐指数（越高越好）：基于真实 Agent 会话，衡量模型是否守规矩——不擅自行动、不虚假归因、不谎报完成。每个信号的失败率先按 1−√失败率换算（让接近满分时的进步依然可见），再加权平均：擅自行动占 50%，虚假归因和谎报完成各占 25%，并已按会话长度校正。数字右下角的小字（例如 ±1.5）是 95% 置信区间。榜单为 Preliminary（初步）状态，数值可能随样本增加调整	擅自行动擅自行动（Unauthorized Action）：Agent 做了超出用户请求或权限范围的操作的会话占比，已按会话长度校正。越低越好，下方小字是 95% 置信区间	虚假归因虚假归因（False Attribution）：Agent 把某句话、请求、选择或事实说成是用户提出的，而用户提供的证据与之矛盾的会话占比，已按会话长度校正。越低越好，下方小字是 95% 置信区间	谎报完成谎报完成（Deceptive Completion）：Agent 明确声称任务已完成，但有具体证据表明并没有完成的会话占比，已按会话长度校正。越低越好，下方小字是 95% 置信区间	开发厂商
1	1 - 4	GPT 6.1 Sol	87.9±1.5	0.89%0.53–1.48%	1.98%1.47–2.68%	2.34%1.77–3.1%	OpenAI

2	1 - 4	GPT 6 Astra	87.8±1.1	0.83%0.56–1.23%	1.96%1.52–2.54%	2.76%2.26–3.37%	OpenAI

3	1 - 4	GPT 6 Luna	87.8±1.4	0.93%0.58–1.48%	1.60%1.14–2.24%	2.90%2.26–3.7%	OpenAI

4	1 - 4	GPT 6 Sol	87.6±1.2	1.16%0.81–1.66%	1.56%1.12–2.17%	2.40%1.9–3.03%	OpenAI

5	5 - 7	GPT 5.6 Sol	84.2±1.2	1.42%1.05–1.9%	2.23%1.74–2.85%	6.06%5.31–6.9%	OpenAI

6	5 - 9	Claude Opus 5.5	83.2±2	1.25%0.68–2.29%	3.86%2.9–5.13%	6.41%5.16–7.94%	Anthropic

7	5 - 8	Grok 4.7	82.7±1.3	1.53%1.08–2.16%	3.03%2.41–3.82%	7.27%6.38–8.26%	SpaceXAI

8	6 - 11	Grok 4.6	80.8±1	1.82%1.43–2.31%	3.61%3.02–4.31%	9.60%8.76–10.5%	SpaceXAI

9	7 - 12	Claude Fable 5.1	80.1±1.2	1.31%0.93–1.85%	5.82%5.01–6.76%	10.50%9.49–11.61%	Anthropic

10	8 - 13	Claude Sonnet 5.5	79.5±1.1	1.83%1.4–2.38%	5.84%5.04–6.76%	9.53%8.54–10.61%	Anthropic

11	8 - 14	Gemini 4 Argon	79.4±1.3	1.30%0.88–1.91%	5.55%4.69–6.56%	12.86%11.69–14.13%	Google

12	9 - 14	Hy4 Preview	78.5±1.1	1.47%1.09–1.98%	6.44%5.62–7.38%	13.24%12.21–14.35%	Tencent

13	10 - 18	Muse Wonder	77.5±1.1	2.09%1.64–2.66%	5.31%4.56–6.16%	14.41%13.37–15.53%	Meta

14	11 - 22	Muse Spark 1.3	77.1±1.1	1.50%1.12–2%	6.95%6.1–7.91%	16.68%15.55–17.87%	Meta

15	13 - 27	Gemini 3.8 Flash	75.8±1.1	2.55%2.02–3.2%	5.99%5.18–6.93%	16.29%15.24–17.4%	Google

16	13 - 28	Claude Sonnet 5	75.7±1.1	1.74%1.33–2.27%	9.77%8.75–10.9%	15.77%14.66–16.94%	Anthropic

17	13 - 28	Claude Opus 5	75.6±1.1	2.32%1.86–2.88%	8.35%7.42–9.39%	14.67%13.65–15.76%	Anthropic

18	13 - 28	R Beam 1.0 501B A23B	75.3±1.6	3.36%2.52–4.46%	4.50%3.58–5.64%	16.84%15.2–18.63%	Reflection AI

19	14 - 28	Step 5 Preview	75.1±1.2	2.11%1.6–2.77%	9.67%8.62–10.84%	15.46%14.33–16.66%	StepFun

20	14 - 28	GLM 5.3	75.1±1	2.16%1.73–2.69%	7.74%6.88–8.69%	17.95%16.85–19.1%	Z.ai

21	14 - 28	Kimi K3	75.0±1.1	2.20%1.74–2.78%	7.61%6.73–8.58%	18.44%17.26–19.68%	Moonshot

22	14 - 28	Deepseek V4.1 Flash	74.8±1.2	2.69%2.13–3.39%	8.92%7.91–10.05%	14.57%13.5–15.72%	DeepSeek

23	15 - 28	Mimo V2.6 Pro	74.5±1.4	2.28%1.7–3.06%	8.32%7.27–9.51%	18.31%16.99–19.72%	Xiaomi

24	15 - 28	Claude Opus 4.8	74.5±1.1	2.32%1.85–2.91%	8.87%7.91–9.94%	17.32%16.12–18.58%	Anthropic

25	15 - 28	Gemini 3.7 Flash	74.4±1	3.45%2.91–4.09%	4.82%4.16–5.59%	18.62%17.58–19.7%	Google