24 models ranked by refusal rate with 95% Wilson confidence intervals. Click any column header to sort. Filter by harm category to see how rankings shift.
#
Model
Refusal Rate
95% CI
Avg Words
Evaluations
1
qwen-2.5-7b-instruct
qwen
91.5%
91.1–91.9%
128
16,135
2
gemini-2.0-flash-lite-001
google
89.9%
89.1–90.6%
199
6,804
3
claude-3-haiku
anthropic
88.2%
87.5–89.0%
207
7,141
4
mistral-large
mistralai
86.1%
84.8–87.3%
210
2,757
5
claude-3.5-haiku
anthropic
84.8%
83.3–86.1%
155
2,596
6
ministral-8b
mistralai
84.7%
83.8–85.5%
142
6,876
7
ministral-14b-2512
mistralai
84.7%
84.2–85.1%
284
23,826
8
qwen-plus
qwen
83.1%
81.7–84.4%
200
2,756
9
grok-3-mini
x-ai
81.4%
80.9–82.0%
193
18,227
10
qwen-2.5-72b-instruct
qwen
79.3%
77.8–80.8%
176
2,724
11
deepseek-chat
deepseek
78.9%
77.5–80.3%
156
3,167
12
mistral-small-24b-instruct-2501
mistralai
76.8%
75.1–78.3%
150
2,757
13
gpt-4o
openai
71.7%
70.1–73.3%
180
3,067
14
gemini-3.1-flash-lite
google
71.5%
70.5–72.6%
195
6,939
15
gemini-2.0-flash-001
google
68.1%
66.3–69.8%
214
2,752
16
gemini-2.5-flash-lite-preview-09-2025
google
67.4%
66.5–68.3%
190
9,910
17
gpt-mini-latest
~openai
67.4%
66.3–68.5%
161
6,939
18
claude-haiku-latest
~anthropic
67.0%
65.8–68.0%
316
6,939
19
gemini-3.1-flash-lite-preview
google
66.7%
65.7–67.7%
193
7,974
20
qwen2.5-coder-7b-instruct
qwen
64.4%
63.6–65.2%
5487
14,031
21
gpt-5.4-mini
openai
64.4%
62.9–65.8%
161
4,010
22
claude-haiku-4.5
anthropic
63.0%
62.3–63.8%
308
16,117
23
claude-3.5-sonnet
anthropic
61.2%
59.4–63.0%
213
2,868
24
gemini-2.5-pro
google
38.4%
34.6–42.3%
237
620
Refusal rate = proportion of prompts where the model refused or added unsolicited caveats. 95% Wilson Score confidence intervals. Higher rank = more restrictive.