LLM Safety and Ethics Index

Which language models have the strongest safety evidence?

An evidence-weighted ranking of 817 language models across 111 independent evaluations. 267 models currently have enough breadth to rank.

What changed · Find evidence · Compare models · Methods · Data

Overall ranking

Rank combines model behavior (90%) and company governance (10%). Models need at least 3 evals across 2 behavior components. The interval is at least as important as the point estimate: it shows source-family sensitivity, not future performance.

How to interpret these ranks

The behavior order is the mean rank across complete-lineage bootstrap refits. Every sub-eval contributes only an ordering; native score scales are never pooled. Intervals separately resample transitive governed-source dependency components. Coverage varies and nearby models are often indistinguishable: the median 90% interval spans 140 ranks.

Rank by release date

257 of 267 ranked models have verified release dates. Each mark identifies the model's developer. Release date is context only and is not used to calculate rank.

What changed

New models and newly ingested eval sources are surfaced here so repeat visitors can quickly see what is new.

Latest models with safety evidence

  1. Qwen3.8 Max · 2 evals across 3 components
  2. Inkling Small · 4 evals across 5 components
  3. Claude Opus 5 · 7 evals across 5 components
  4. Gemini 3.6 Flash · 5 evals across 5 components
  5. Gemini 3.5 Flash Lite · 6 evals across 4 components
  6. G9V3 3B · 1 evals across 1 components
  7. Kimi K3 · 13 evals across 6 components
  8. Inkling · 7 evals across 6 components

Recently ingested evals

  1. OpenAI o3-mini System Card 2 sub-evals · 2 models
  2. OpenAI o3 and o4-mini System Card 2 sub-evals · 2 models
  3. OpenAI GPT-5 System Card 1 sub-evals · 2 models
  4. OpenAI GPT-4o System Card 1 sub-evals · 2 models
  5. Google Gemini 2.5 Flash-Lite Model Card 1 sub-evals · 2 models

Full ranking table

Rows are ordered by estimated rank, but broad overlapping intervals should not be read as evidence of a real difference.

267 models
Overall rank; brackets show the 90% source-sensitivity interval.
Estimated rank
[90% interval]		ModelEvalsComponentsRelease date
1 [1–6]claude-opus-575/72026-07-24
2 [2–12]claude-fable-5167/72026-06-09
3 [1–36]claude-264/72023-07-11
4 [3–49]claude-opus-4.8207/72026-05-28
5 [3–50]claude-opus-4.7257/72026-04-16
6 [3–56]claude-opus-4197/72025-05-22
7 [5–60]gpt-5.6-terra177/72026-07-09
8 [7–51]claude-opus-4.6237/72026-02-05
9 [10–66]gpt-5.5297/72026-04-23
10 [5–186]muse-spark-1.1117/72026-07-09
11 [5–55]gpt-5-nano18+27/72025-08-07
12 [3–86]gpt-5.3-chat55/72026-03-03
13 [3–132]gpt-5.2-chat33/72025-12-11
14 [4–72]claude-sonnet-4.6257/72026-01-21
15 [3–135]gpt-5.5-instant33/72026-05-26
16 [13–74]gpt-5.2267/72025-12-11
17 [11–71]claude-haiku-4.5327/72025-10-15
18 [6–94]gemini-3.1-pro-preview207/72026-02-19
19 [4–142]gemini-3.6-flash55/72026-07-21
20 [13–96]gpt-5-mini23+17/72025-08-07
21 [10–79]claude-3.7-sonnet257/72025-02-24
22 [8–123]muse-spark33/72026-04-08
23 [5–111]claude-opus-4.5207/72025-11-24
24 [11–100]gpt-5.4227/72026-03-05
25 [13–83]claude-sonnet-4317/72025-05-22
26 [12–87]claude-sonnet-4.5327/72025-09-29
27 [10–81]claude-opus-4.1126/72025-08-05
28 [16–128]qwen3.5-27b36/72026-02-24
29 [11–162]grok-4.20-multi-agent44/72026-03-10
30 [13–148]inkling-small2+25/72026-07-30
31 [8–158]phi-3.5-moe-instruct56/72024-08-22
32 [6–175]gpt-5.6-sol177/72026-07-09
33 [6–132]o1-preview46/72024-09-12
34 [8–162]inkling5+26/72026-07-14
35 [4–179]gemini-2.5-pro-exp43/72025-03-25
36 [2–232]gpt-5-pro34/72025-08-07
37 [13–157]minimax-m35+15/72026-06-02
38 [14–135]gpt-4.5-preview96/72025-02-27
39 [6–178]mimo-v2.5-pro4+15/72026-04-27
40 [9–207]llama-3.1-tulu-3-8b35/72024-11-20
41 [15–139]gpt-5327/72025-08-07
42 [33–133]qwen3.5-397b-a17b7+17/72026-02-16
43 [18–256]palmyra-x-00463/72024-10-09
44 [22–183]glm-4.5-air10+15/72025-07-28
45 [22–152]kimi-k3136/72026-07-16
46 [5–150]claude-3.5-haiku147/72024-10-22
47 [29–171]glm-59+17/72026-02-11
48 [39–184]deepseek-v4-flash13+17/72026-04-22
49 [12–180]nemotron-3-ultra-550b-a55b4+13/72026-06-03
50 [20–179]nemotron-3-super-120b-a12b6+15/72026-03-10
51 [31–139]qwen2-72b-instruct105/72024-06-07
52 [14–221]internlm2-chat-7b44/72024-01-17
53 [39–144]gpt-5.1267/72025-11-13
54 [27–251]qwen3-235b-a22b-fp8-throughput63/72025-04-28
55 [20–136]gpt-4.1257/72025-04-14
56 [8–224]nova-lite56/72024-12-03
57 [22–183]gemini-1.0-pro86/72023-12-06
58 [31–187]qwen2.5-32b-instruct45/72024-09-19
59 [21–222]internlm2-chat-20b45/72024-01-17
60 [28–180]qwen-72b-chat44/72023-11-30
61 [62–206]palmyra-fin63/72024-07-31
62 [11–149]o3236/72025-04-16
63 [37–151]kimi-k2.616+27/72026-04-14
64 [14–214]qwen3.6-max-preview33/72026-04-22
65 [27–190]grok-4.5136/72026-07-08
66 [22–173]gemini-3.5-flash207/72026-05-19
67 [20–185]internlm-chat-20b44/72023-09-18
68 [32–195]grok-4.20187/72026-03-10
69 [14–182]gpt-5.6-luna16+17/72026-07-09
70 [29–160]gemini-1.5-pro207/72024-02-15
71 [3–185]claude-sonnet-5107/72026-06-30
72 [27–190]palmyra-x563/72025-04-28
73 [25–200]gpt-5.2-codex33/7—
74 [48–160]kimi-k225+17/72025-07-11
75 [20–206]qwen3.5-9b3+15/72026-02-27
76 [51–148]gpt-oss-120b22+37/72025-08-05
77 [28–224]mistral-large66/72024-02-26
78 [21–156]o1166/72024-12-17
79 [46–151]deepseek-v3.222+17/72025-12-01
80 [15–198]qwen3.7-max76/72026-05-20
81 [39–180]gpt-oss-20b15+17/72025-08-05
82 [30–220]doubao-seed-1.645/72025-06-11
83 [20–160]gemini-3-pro-preview227/72025-11-18
84 [29–177]qwen3-max66/72025-09-24
85 [52–167]kimi-k2.520+27/72026-01-01
86 [45–188]glm-5.214+16/72026-06-16
87 [28–184]grok-3-mini157/72025-04-03
88 [35–191]mimo-v2.55+15/72026-04-27
89 [17–196]qwen3.6-plus66/72026-04-02
90 [15–207]qwen3-coder-480b-a35b-instruct32/72025-07-22
91 [27–180]gemini-3.5-flash-lite5+14/72026-07-21
92 [18–199]o1-mini117/72024-09-12
93 [36–167]gemini-1.5-flash156/72024-05-14
94 [50–191]minimax-m2.710+16/72026-04-09
95 [51–191]grok-4.1-fast167/72025-11-19
96 [42–211]gpt-5.3-codex55/72026-02-05
97 [27–182]gpt-oss-safeguard-20b3+15/72025-10-29
98 [53–182]grok-4.3197/72026-05-15
99 [46–173]claude-3-opus227/72024-03-04
100 [41–173]qwen3-235b-a22b25+27/72025-04-29
101 [14–178]gpt-4.1-mini16+17/72025-04-14
102 [40–174]grok-4-fast125/72025-09-19
103 [41–196]gpt-4147/72023-03-14
104 [55–173]gemini-2.5-pro337/72025-03-25
105 [47–164]grok-4277/72025-07-09
106 [44–192]glm-5.114+26/72026-04-03
107 [40–230]deepseek-v4-pro127/72026-04-22
108 [70–202]qwen3-next-80b-a3b106/72025-09-11
109 [55–160]claude-3.5-sonnet277/72024-06-21
110 [33–161]gpt-4-turbo197/72023-11-06
111 [12–216]qwen3.7-plus45/72026-05-26
112 [42–186]qwen-14b-chat84/72023-09-25
113 [13–230]phi-455/72024-12-11
114 [19–217]qwen2.5-72b-instruct15+17/72024-09-16
115 [50–194]gemma-4-31b-it7+17/72026-03-11
116 [50–203]gemma-4-26b-a4b-it4+15/72026-03-11
117 [21–245]phi-3-mini-4k-instruct43/72024-04-22
118 [36–220]deepseek-v3.110+17/72025-08-21
119 [27–229]minimax-m256/72025-10-27
120 [57–178]deepseek-v329+17/72024-12-26
121 [44–186]gemini-2.5-flash30+17/72025-04-17
122 [43–230]qwen3-30b-a3b-instruct46/72025-07-28
123 [63–223]glm-4.656/72025-09-30
124 [12–215]granite-4.0-small-guardian63/72025-10-02
125 [55–186]gemini-3.1-flash-lite18+27/72026-03-03
126 [58–254]mistral-medium45/72023-12-11
127 [26–208]ernie-4.5-300b-a47b56/72025-06-28
128 [84–166]gpt-4o55+17/72024-05-13
129 [58–187]qwen1.5-32b-chat46/72024-04-02
130 [78–194]gemini-2-flash75/72024-12-11
131 [80–201]llama-3.3-70b-instruct22+27/72024-12-06
132 [26–249]glm-4.7-flash3+13/72026-01-19
133 [45–264]qwen3-30b-a3b45/72025-04-27
134 [58–181]gpt-3.5-turbo237/72023-03-01
135 [65–184]baichuan-2-13b-chat125/72023-09-06
136 [79–192]deepseek-r1397/72025-01-20
137 [1–232]claude-2.132/72023-11-21
138 [59–254]glm-4.765/72026-06-03
139 [58–188]qwen1.5-72b-chat126/72024-02-04
140 [64–201]llama-2-7b-chat127/72023-07-18
141 [57–198]qwen1.5-14b-chat65/72024-02-04
142 [75–176]o4-mini246/72025-04-16
143 [23–250]qwen3-32b5+17/72025-04-29
144 [69–218]minimax-m2.58+17/72026-02-12
145 [20–224]nova-premier45/7—
146 [69–204]gemini-3-flash-preview14+16/72025-12-17
147 [9–242]hy32+15/72026-07-02
148 [39–215]internlm-chat-7b54/72023-07-06
149 [18–236]deepseek-v3.2-exp34/72025-09-29
150 [58–199]yi-34b-chat85/72023-11-23
151 [60–207]text-davinci-00334/72022-11-28
152 [69–235]glm-4-9b-chat74/72024-06-05
153 [69–181]gemini-2-pro-preview63/72025-02-05
154 [67–245]magistral-medium66/72025-06-10
155 [66–201]command-r-plus107/72024-04-04
156 [25–240]nemotron-3-nano-30b-a3b4+14/72025-12-15
157 [45–222]mimo-v2-flash45/72025-12-16
158 [42–202]gemma-2b-it56/72024-02-21
159 [84–200]o3-mini236/72025-01-31
160 [59–205]granite-4.0-small63/72025-10-02
161 [68–197]claude-3-haiku177/72024-03-13
162 [65–201]gemini-2.0-flash126/72024-12-11
163 [52–228]grok-3177/72025-04-03
164 [61–214]llama-2-13b-chat86/72023-07-18
165 [104–201]llama-4-maverick27+17/72025-04-05
166 [96–205]llama-3.1-405b-instruct226/72024-07-23
167 [48–208]gemma-7b-it77/72024-02-21
168 [54–218]gpt-5.4-mini146/72026-03-17
169 [86–203]qwen1.5-7b-chat96/72024-02-04
170 [68–219]glm-4.597/72025-07-28
171 [65–251]mistral-small-3.142/72025-03-11
172 [89–203]yi-6b-chat55/72023-11-23
173 [46–213]gemma-2-9b-it77/72024-06-27
174 [100–232]gpt-4.1-nano126/72025-04-14
175 [39–255]qwen3.5-35b-a3b34/72026-02-24
176 [25–248]hy3-preview43/72026-04-13
177 [28–237]minimax-m2.143/72025-12-20
178 [65–236]llama-3.2-instruct-11b-vision34/7—
179 [77–201]gemini-2-flash-lite-preview63/72025-02-05
180 [15–226]granite-4.0-micro-guardian63/72025-10-02
181 [72–247]glm-5-turbo45/72026-03-15
182 [108–218]gemma-2-27b-it57/72024-06-27
183 [93–213]llama-3-8b-instruct157/72024-04-18
184 [29–259]deepseek-llm-7b-chat33/72023-11-29
185 [93–220]llama-3-70b-instruct137/72024-04-18
186 [48–258]command-r84/72024-03-11
187 [53–213]llama-3.1-70b-instruct197/72024-07-23
188 [49–234]olmo-2-0325-32b-instruct76/72025-03-13
189 [120–217]vicuna-13b33/72023-03-30
190 [110–214]qwen-7b-chat94/72023-08-03
191 [119–217]baichuan-2-7b-chat115/72023-09-06
192 [63–233]llama-2-70b-chat76/72023-07-18
193 [73–234]phi-3.5-mini-instruct56/72024-08-22
194 [48–233]gemini-2.5-flash-lite187/72025-06-17
195 [79–261]phi-233/7—
196 [142–205]claude-3-sonnet136/72024-03-04
197 [85–235]gpt-5.4-nano125/72026-03-17
198 [75–247]qwen-2.5-7b-instruct15+17/72024-09-19
199 [102–234]chatglm-6b54/72023-03-14
200 [14–255]nvidia-nemotron-nano-9b-v232/72025-08-12
201 [103–227]qwen-2.5-14b-instruct65/72024-09-19
202 [127–222]chatglm3-6b94/72023-10-27
203 [93–249]step-3.5-flash3+15/72026-02-01
204 [8–247]gemini-2.0-flash-lite44/72025-02-25
205 [139–236]mistral-7b-instruct217/72023-09-27
206 [17–244]yi-1.5-34b-chat34/72024-05-13
207 [70–254]mistral-small-3.235/72025-06-19
208 [132–228]vicuna-7b-v1.544/72023-06-22
209 [116–246]mistral-medium-3.156/72025-08-12
210 [60–245]nova-micro45/7—
211 [148–221]gpt-4o-mini247/72024-07-18
212 [29–254]granite-3.3-8b-instruct84/72025-04-10
213 [56–245]gemma-3-27b-it106/72025-03-12
214 [100–265]mistral-nemo45/72024-07-18
215 [102–264]mistral-medium-343/72025-05-07
216 [164–223]llama-4-scout167/72025-04-05
217 [132–244]mistral-small-444/72026-01-23
218 [103–242]phi-4-mini34/7—
219 [141–237]llama-3.1-8b-instruct257/72024-07-23
220 [59–248]gemma-2-2b-it45/72024-07-31
221 [147–248]qwen2.5-max46/72025-01-25
222 [146–237]vicuna-13b-v1.544/72023-03-30
223 [170–241]palmyra-med63/72024-07-31
224 [95–260]qwen-2.5-1.5b-instruct53/72024-09-19
225 [157–250]chatglm2-6b63/72023-06-25
226 [54–265]qwen2.5-3b-instruct33/72024-09-17
227 [42–257]qwen1.5-0.5b-chat34/72024-02-04
228 [144–253]llama-3.2-90b-vision-instruct33/72024-09-19
229 [134–249]nova-pro66/72024-12-03
230 [77–262]qwen2-7b-instruct43/72024-06-07
231 [25–258]olmo-2-1124-13b-instruct65/72024-11-26
232 [102–262]ling-2.6-flash3+15/7—
233 [96–257]zephyr-7b-beta54/72023-10-26
234 [72–237]granite-4.0-micro96/72025-10-02
235 [110–263]mistral-medium-3.534/72026-03-31
236 [176–246]vicuna-7b43/72023-06-22
237 [48–260]mixtral-8x22b-instruct95/72024-04-17
238 [159–255]qwen3-8b66/72025-04-29
239 [180–264]mistral-small-386/72025-01-30
240 [163–250]trinity-large44/72026-01-27
241 [183–258]qwen3-14b36/72025-04-27
242 [183–255]moss-moon-003-sft33/72023-04-19
243 [195–253]gemma-3-4b34/7—
244 [200–257]command-a65/72025-03-11
245 [192–256]vicuna-33b-v1.344/72023-06-22
246 [188–256]davinci34/72020-06-11
247 [140–262]mistral-large-396/72025-12-01
248 [206–253]gemma-3-12b45/7—
249 [131–265]deepseek-v3.2-speciale33/72025-11-28
250 [189–258]llama-3.2-3b-instruct86/72024-09-25
251 [120–267]deepseek-llm-67b-chat85/72023-11-29
252 [133–265]magistral-small35/72025-06-04
253 [191–266]llama-3.2-1b-instruct76/72024-09-25
254 [158–258]marin-8b-instruct53/72025-05-14
255 [180–264]qwq-32b45/72025-03-06
256 [224–262]jamba-1-7-large34/7—
257 [200–263]qwen-1.8b-chat33/72023-11-30
258 [196–262]olmo-2-1124-7b-instruct65/72024-11-26
259 [173–267]grok-243/72024-12-12
260 [188–265]mixtral-8x7b-instruct106/72023-12-10
261 [236–263]dbrx-instruct75/72024-03-27
262 [228–264]mistral-large-2115/72024-07-24
263 [178–265]olmoe-1b-7b-0125-instruct53/72025-01-27
264 [229–266]qwen1.5-1.8b-chat45/72024-02-04
265 [235–266]deepseek-r1-distill-llama-70b43/72025-01-20
266 [246–267]qwen-2.5-0.5b-instruct43/72024-09-19
267 [238–267]magistral-small-250934/72025-09-12