Comparing language models on safety and ethical behavior

The index combines results from independent evaluations of harmful assistance, truthfulness, rights and welfare, moral consideration, and responsible use of power. Company governance contributes 10% of the overall score.

400
models analyzed
124
models with enough evidence to rank
53
evals
137
sub-evals

How to read the ranking

Scores are relative summaries of the available safety evidence. Coverage differs across models, and nearby ranks are often indistinguishable: the median 90% rank interval spans 56 places. The most obvious current issue is old models, which were good relative to the models of their day, ranking highly compared to modern, safer models, because of insufficient overlap in evals between old and new models.

Overall ranking

Models appear only after they have results from at least 3 evals covering at least 2 of the five behavior components. Note that e.g. Kimi K3 (capable and new) is missing due to a lack of sufficient evals.

124 models
Overall scores combine model behavior (90%) and company governance (10%). Brackets are 90% uncertainty intervals.
RankModelOverall scoreGovernanceEvalsComponentsRelease date
1 [1–37]Anthropicclaude-opus-4.796.7 [57.2–98.8]87.4Direct135/52026-04-16
2 [1–11]Anthropicclaude-3.7-sonnet94.3 [75.9–97.0]87.3Direct104/52025-02-24
3 [1–15]Anthropicclaude-opus-4.893.5 [71.2–96.9]87.4Direct105/52026-05-28
4 [1–25]Anthropicclaude-3.5-sonnet90.9 [64.9–96.8]87.3Direct145/52024-06-21
5 [2–24]OpenAIgpt-5.590.2 [66.6–93.9]82.4Direct145/52026-04-23
6 [2–20]OpenAIgpt-488.1 [67.3–94.5]62.0Direct103/52023-03-14
7 [2–41]Anthropicclaude-sonnet-4.687.2 [57.5–94.1]87.6Direct135/52026-01-21
8 [4–51]Anthropicclaude-3-opus80.6 [52.6–89.9]87.3Direct135/52024-03-04
9 [3–53]Googlegemini-1.5-pro79.9 [52.9–91.0]68.3Direct135/52024-02-15
10 [5–33]Metamuse-spark-1.179.2 [60.3–88.3]59.6Direct74/52026-07-09
11 [5–38]Anthropicclaude-fable-579.1 [57.4–89.1]87.4Direct74/52026-06-09
12 [3–94]Mistral AImistral-large74.5 [39.6–91.8]26.0Direct35/52024-02-26
13 [7–35]Anthropicclaude-opus-4.673.8 [56.6–83.2]87.6Direct115/52026-02-05
14 [10–39]Anthropicclaude-opus-4.571.9 [56.6–80.6]87.6Direct84/52025-11-24
15 [4–101]Anthropicclaude-sonnet-4.571.8 [36.3–90.5]87.6Direct145/52025-09-29
16 [10–46]Googlegemini-1.5-flash71.5 [54.8–81.4]68.3Direct105/52024-05-14
17 [9–49]OpenAIgpt-4-turbo68.7 [53.3–80.7]62.0Direct114/52023-11-06
18 [7–72]internlm-chat-7b66.8 [47.4–83.4]49.8Median-imputed32/52023-07-06
19 [9–78]OpenAIgpt-5.266.7 [45.1–79.3]81.7Direct125/52025-12-11
20 [6–105]Metallama-2-7b-chat66.5 [34.1–84.5]25.3Direct92/52023-07-18
21 [8–93]Metallama-3.1-405b-instruct66.4 [40.2–83.1]25.3Direct104/52024-07-23
22 [10–56]Anthropicclaude-3-haiku66.4 [51.4–79.0]87.3Direct103/52024-03-13
23 [9–63]OpenAIgpt-5-nano64.1 [49.9–79.7]73.8Direct114/52025-08-07
24 [15–52]Alibabaqwen-14b-chat63.7 [52.6–72.2]29.5Direct52/52023-09-25
25 [9–80]yi-34b-chat63.4 [44.2–80.4]49.8Median-imputed62/52023-11-23
26 [13–62]OpenAIgpt-5-mini62.8 [50.1–74.9]73.8Direct125/52025-08-07
27 [9–87]Anthropicclaude-3.5-haiku62.4 [42.8–80.8]87.3Direct45/52024-10-22
28 [6–105]Googlegemini-2.5-flash62.3 [35.1–84.3]64.9Direct155/52025-04-17
29 [15–70]Z.aiglm-562.1 [47.6–72.9]38.3Direct34/52026-02-12
30 [13–91]OpenAIo360.7 [41.4–76.2]73.8Direct104/52025-04-16
31 [13–96]OpenAIgpt-5.6-sol60.3 [39.3–74.7]82.4Direct74/52026-07-09
32 [10–104]Alibabaqwen3-235b-a22b60.1 [34.8–79.6]29.5Direct105/52025-04-29
33 [11–92]OpenAIgpt-oss-120b60.1 [40.9–78.1]73.8Direct95/52025-08-05
34 [21–66]OpenAIgpt-5.6-terra59.6 [49.1–68.2]82.4Direct74/52026-07-09
35 [13–81]Anthropicclaude-sonnet-459.6 [44.6–76.5]88.2Direct114/52025-05-22
36 [22–50]OpenAIgpt-4.159.2 [53.3–65.8]73.8Direct83/52025-04-14
37 [26–47]Anthropicclaude-opus-458.5 [54.3–63.0]88.2Direct52/52025-05-22
38 [7–111]OpenAIgpt-3.5-turbo58.3 [28.5–82.5]62.0Direct175/52023-03-01
39 [17–95]Z.aiglm-5.157.1 [39.0–70.0]37.8Direct84/52026-04-07
40 [28–57]OpenAIgpt-5.157.0 [51.3–62.2]81.7Direct124/52025-11-13
41 [15–104]Anthropicclaude-haiku-4.556.7 [35.0–72.5]87.6Direct165/52025-10-15
42 [21–79]xAIgrok-4-255.6 [45.6–66.9]37.4Direct74/52026-03-10
43 [25–85]internlm2-chat-7b55.2 [44.1–63.9]49.8Median-imputed42/52024-01-17
44 [25–73]internlm2-chat-20b55.1 [47.0–64.1]49.8Median-imputed32/52024-01-17
45 [23–89]baichuan-2-13b-chat54.1 [42.3–67.3]49.8Median-imputed103/52023-09-06
46 [25–82]OpenAIo4-mini53.8 [45.0–65.0]73.8Direct104/52025-04-16
47 [29–82]OpenAIo153.5 [45.1–63.2]62.0Direct83/52024-12-17
48 [11–109]Metallama-3.3-70b-instruct53.2 [31.0–77.9]25.3Direct84/52024-12-06
49 [15–102]Moonshot AIkimi-k253.1 [35.7–73.4]49.8Median-imputed124/52025-07-11
50 [31–79]OpenAIgpt-5.6-luna53.0 [44.9–60.7]82.4Direct74/52026-07-09
51 [38–64]Googlegemini-1.0-pro52.9 [49.5–57.5]68.3Direct62/52023-12-06
52 [33–73]OpenAIo1-mini52.1 [46.8–60.5]62.0Direct62/52024-09-12
53 [34–70]Ai2olmo-2-0325-32b-instruct52.1 [48.0–58.4]49.8Median-imputed44/52025-03-13
54 [39–71]Googlegemma-7b-it52.0 [47.6–57.0]68.3Direct42/52024-02-21
55 [16–111]DeepSeekdeepseek-v351.9 [29.3–72.6]21.9Direct164/52024-12-26
56 [44–63]Googlegemini-2-flash51.6 [49.9–55.3]68.3Direct54/52024-12-11
57 [44–67]Googlegemma-2-27b-it51.5 [48.7–55.7]68.3Direct32/52024-06-27
58 [25–101]OpenAIgpt-5.451.4 [36.7–64.1]82.4Direct84/52026-03-05
59 [30–90]MiniMaxminimax-m2.751.1 [43.1–61.7]49.8Median-imputed34/52026-04-09
60 [48–66]OpenAIgpt-4.5-preview50.9 [49.2–54.3]62.0Direct52/52025-02-27
61 [23–103]Metallama-3-8b-instruct50.3 [36.2–66.1]25.3Direct92/52024-04-18
62 [22–107]xAIgrok-4-fast50.0 [32.4–67.9]44.0Direct85/52025-09-19
63 [41–84]OpenAIgpt-oss-20b50.0 [43.6–56.8]73.8Direct64/52025-08-05
64 [30–90]yi-1.5-34b-chat49.6 [41.4–61.6]49.8Median-imputed34/52024-05-13
65 [13–110]DeepSeekdeepseek-r149.4 [31.4–74.3]21.9Direct185/52025-01-20
66 [24–107]Anthropicclaude-3-sonnet48.0 [32.6–66.3]87.3Direct93/52024-03-04
67 [41–95]Alibabaqwen-7b-chat47.8 [39.2–56.8]29.5Direct73/52023-08-03
68 [53–81]Googlegemma-2b-it47.6 [44.4–52.7]68.3Direct32/52024-02-21
69 [35–103]Googlegemma-2-9b-it47.0 [36.0–58.5]68.3Direct64/52024-06-27
70 [48–92]Ai2olmo-2-1124-13b-instruct46.5 [40.8–54.1]49.8Median-imputed44/52024-11-26
71 [18–114]Googlegemini-3.1-pro-preview46.3 [28.2–70.6]72.3Direct104/52026-02-19
72 [38–97]xAIgrok-4.545.8 [38.5–57.4]37.4Direct74/52026-07-08
73 [38–101]Alibabaqwen1.5-7b-chat45.7 [36.3–57.4]29.5Direct82/52024-02-04
74 [25–106]Alibabaqwen-2.5-72b-instruct45.3 [32.4–63.2]29.5Direct85/52024-09-19
75 [24–108]xAIgrok-4.1-fast45.2 [32.1–65.9]44.0Direct95/52025-11-19
76 [52–95]OpenAIgpt-4.1-mini45.1 [40.0–52.6]73.8Direct63/52025-04-14
77 [49–98]yi-6b-chat44.9 [38.7–53.9]49.8Median-imputed52/52023-11-23
78 [72–88]vicuna-7b-v1.544.0 [42.5–47.2]25.3Base-model inherited42/52023-06-22
79 [59–98]Coherecommand-r43.3 [37.1–50.9]49.8Median-imputed53/52024-03-11
80 [39–107]DeepSeekdeepseek-v4-flash42.9 [32.4–57.0]49.8Median-imputed44/5
81 [35–109]Googlegemini-3-pro-preview42.7 [31.5–58.6]72.3Direct103/52025-11-18
82 [33–108]DeepSeekdeepseek-v4-pro42.6 [31.6–58.9]49.8Median-imputed44/5
83 [8–120]OpenAIgpt-542.5 [14.7–81.3]73.8Direct154/52025-08-07
84 [46–106]xAIgrok-442.4 [32.8–54.2]50.0Direct154/52025-07-09
85 [45–110]OpenAIgpt-5.4-nano40.9 [29.2–54.5]82.4Direct74/52026-03-17
86 [49–111]Z.aiglm-5.240.5 [29.2–53.4]37.8Direct95/52026-06-16
87 [73–99]Metallama-4-scout40.4 [36.8–47.4]41.7Direct64/52025-04-05
88 [31–112]Z.aiglm-4-9b-chat40.2 [28.1–61.1]49.5Direct64/52024-06-05
89 [22–116]OpenAIo3-mini40.0 [22.2–67.0]62.0Direct114/52025-01-31
90 [71–103]Coherecommand-r-plus39.8 [34.9–47.3]49.8Median-imputed63/52024-04-04
91 [74–103]vicuna-13b-v1.539.7 [35.6–46.4]25.3Base-model inherited42/52023-03-30
92 [36–112]OpenAIgpt-5.4-mini39.6 [26.7–58.0]82.4Direct94/52026-03-17
93 [62–109]Metallama-2-13b-chat38.8 [31.2–50.0]25.3Direct73/52023-07-18
94 [68–108]Moonshot AIkimi-k2.6-64k37.9 [31.8–49.0]49.8Median-imputed74/52026-04-14
95 [79–108]Mistral AImixtral-8x7b-instruct-v0.137.4 [31.9–45.7]26.0Direct72/52023-12-11
96 [68–112]xAIgrok-337.1 [28.3–48.6]50.0Direct62/52025-04-03
97 [68–111]Moonshot AIkimi-k2.636.7 [28.1–48.4]49.8Median-imputed34/52026-04-14
98 [62–112]Moonshot AIkimi-k2.536.6 [29.4–50.1]49.8Median-imputed84/52026-01-01
99 [19–121]Z.aichatglm-6b34.8 [13.1–66.8]49.5Direct42/52023-03-14
100 [25–119]OpenAIgpt-4o34.6 [18.1–63.1]62.0Direct275/52024-05-13
101 [83–112]zephyr-7b-beta34.6 [28.3–44.1]26.0Base-model inherited32/52023-10-26
102 [62–116]Googlegemini-3.5-flash34.2 [24.1–50.1]75.8Direct115/52026-05-19
103 [75–113]Googlegemini-3-flash-preview33.3 [27.0–46.7]72.3Direct74/52025-12-17
104 [58–116]Alibabaqwen-2.5-14b-instruct32.0 [22.2–51.3]29.5Direct44/52024-09-19
105 [60–118]DeepSeekdeepseek-v3.231.5 [20.5–50.8]29.5Direct125/52025-12-01
106 [48–119]Metallama-3-70b-instruct31.3 [17.5–54.1]25.3Direct83/52024-04-18
107 [16–122]Googlegemini-2.5-pro30.6 [12.0–72.5]64.9Direct174/52025-03-25
108 [35–122]Metallama-4-maverick30.4 [12.0–59.3]41.7Direct105/52025-04-05
109 [66–118]Z.aichatglm3-6b30.1 [17.3–49.0]49.5Direct62/52023-10-27
110 [84–119]Mistral AImistral-large-328.0 [18.7–44.3]26.0Direct34/52025-12-01
111 [93–118]Mistral AImistral-7b-instruct26.2 [18.6–41.0]26.0Direct123/52023-09-27
112 [51–122]baichuan-2-7b-chat24.3 [10.8–53.0]49.8Median-imputed82/52023-09-06
113 [55–121]Googlegemini-3.1-flash-lite22.8 [12.8–53.1]75.8Direct105/52026-03-03
114 [31–123]Metallama-3.1-70b-instruct21.2 [5.6–61.1]25.3Direct114/52024-07-23
115 [42–123]Alibabaqwen-2.5-7b-instruct20.4 [8.1–55.3]29.5Direct105/52024-09-19
116 [51–124]DeepSeekdeepseek-v3.119.5 [4.9–53.5]21.9Direct43/52025-08-21
117 [21–124]xAIgrok-4.2018.7 [4.3–68.6]37.4Direct34/52026-03-10
118 [105–122]Googlegemini-2.5-flash-lite17.7 [11.8–33.5]64.9Direct125/52025-06-17
119 [76–123]xAIgrok-3-mini17.7 [7.2–46.1]50.0Direct74/52025-04-03
120 [54–123]OpenAIgpt-4o-mini16.7 [8.4–53.2]62.0Direct125/52024-07-18
121 [53–124]xAIgrok-4.312.5 [6.9–52.1]37.4Direct105/52026-04-17
122 [95–124]Mistral AImistral-large-211.8 [4.4–40.3]26.0Direct54/52024-07-24
123 [85–124]Z.aichatglm2-6b9.4 [2.3–43.2]49.5Direct53/52023-06-25
124 [77–124]Metallama-3.1-8b-instruct9.0 [2.3–46.3]25.3Direct124/52024-07-23

Score by release date

Overall score by release date, with each model shown by its developer's mark.
122 of 124 ranked models have verified release dates. Each mark identifies the model's developer. Release date is shown for context and is not used to calculate the score.