Evaluation profile

LiveSecBench

5sub-evals

1.28%total index weight

4components

Within-component eval weight: Fairness 2.23% · Truthfulness 1.68% · Misuse resistance 0.749% · Human rights 4.9%.

Model score (higher is better)Predicted score

About this eval

Live security benchmark performance for Chinese and international models.

Included in the behavior ranking.

Sub-evals

Measure	Component	Direction	Total index weight	Within-component weight
ethicslivesecbench/livesecbench.csv:ethicsMeasures whether the model avoids unfair, biased, or unethical content across continuously updated safety prompts.	fairness_nondiscrimination:1.000livesecbench/livesecbench.csv	Higher is better	0.223%	Fairness 2.23%
factualitylivesecbench/livesecbench.csv:factualityMeasures whether the model avoids false or misleading claims across continuously updated safety prompts.	truthfulness_honesty:1.000livesecbench/livesecbench.csv	Higher is better	0.252%	Truthfulness 1.68%
legalitylivesecbench/livesecbench.csv:legalityMeasures whether the model avoids illegal or public-safety harms across continuously updated safety prompts.	ordinary_harm_misuse_resistance:1.000livesecbench/livesecbench.csv	Higher is better	0.0749%	Misuse resistance 0.749%
privacylivesecbench/livesecbench.csv:privacyMeasures whether the model avoids privacy violations or disclosure of personal information across continuously updated safety prompts.	human_rights_systemic_harm:1.000livesecbench/livesecbench.csv	Higher is better	0.459%	Human rights 3.06%
psychological_healthlivesecbench/livesecbench.csv:psychological_healthMeasures whether the model avoids responses that could harm a user’s psychological health across continuously updated safety prompts.	human_rights_systemic_harm:1.000livesecbench/livesecbench.csv	Higher is better	0.276%	Human rights 1.84%

ethics

Measures whether the model avoids unfair, biased, or unethical content across continuously updated safety prompts.

Rank	Model	Value	Provenance
1	gpt-5.2	95.79	official
2	qwen3.5-397b-a17b	89.63	official
3	spark-x2	89.26	official
4	claude-haiku-4.5	88.65	official
5	qwen3.5-plus	84.35	official
6	mimo-v2-flash	80.99	official
7	kimi-k2.5	79.3	official
8	claude-sonnet-4.6	78.63	official
9	deepseek-v3.2	75.93	official
10	minimax-m2.5	72.04	official
11	doubao-seed-1.6	71.6	official
12	minimax-m2	69.21	official
13	intern-s1-pro	65.65	official
14	gpt-oss-120b	61.84	official
15	qwen3-235b-a22b	60.81	official
16	step-3.5-flash	56.5	official
17	hunyuan-t1	55.14	official
18	doubao-seed-2.0-pro	55.1	official
19	llama-3.3-70b-instruct	54.35	official
20	sensenova-v6.5-turbo	51.6	official
21	sensechat-turbo	47.69	official
22	seed-oss-36b-instruct	46.86	official
23	gemini-3.1-pro-preview	45.45	official
24	longcat-flash-chat	45.19	official
25	step-3	42.42	official
26	deepseek-r1	39.41	official
27	glm-4.6	36.13	official
28	intern-s1	34.35	official
29	ernie-5.0-preview	34.13	official
30	ernie-4.5-21b-a3b	33.56	official
31	gemini-2.5-flash	31.43	official
32	glm-5	31.18	official
33	mistral-large-2	27.9	official
34	mistral-large-3	24.59	official
35	gpt-4.1-mini	23.73	official
36	ling-2.5-1t	23.32	official
37	llama-4-maverick	22.55	official
38	deepseek-v3	20.88	official
39	grok-3-mini	20.84	official
40	kimi-k2	20.12	official
41	hunyuan-a13b-instruct	14.7	official
42	pangu-pro-moe-72b-a16b	9.51	official
43	grok-4.1-fast	8.57	official

factuality

Measures whether the model avoids false or misleading claims across continuously updated safety prompts.

Rank	Model	Value	Provenance
1	claude-sonnet-4.6	92.17	official
2	qwen3.5-plus	92.1	official
3	claude-haiku-4.5	82.68	official
4	kimi-k2.5	81.18	official
5	minimax-m2.5	76.82	official
6	glm-5	75.94	official
7	spark-x2	73.79	official
8	qwen3.5-397b-a17b	72.1	official
9	qwen3-235b-a22b	69.98	official
10	longcat-flash-chat	68.41	official
11	doubao-seed-2.0-pro	67.71	official
12	step-3.5-flash	62.37	official
13	minimax-m2	62.29	official
14	glm-4.6	58.59	official
15	intern-s1-pro	58.23	official
16	gpt-5.2	57.75	official
17	ling-2.5-1t	55.88	official
18	doubao-seed-1.6	55.43	official
19	mimo-v2-flash	52.77	official
20	grok-4.1-fast	52.76	official
21	gemini-3.1-pro-preview	51.4	official
22	gpt-oss-120b	48.37	official
23	step-3	46.32	official
24	mistral-large-3	43.2	official
25	intern-s1	41.43	official
26	deepseek-v3.2	40.76	official
27	gemini-2.5-flash	38.55	official
28	hunyuan-a13b-instruct	37.38	official
29	grok-3-mini	37.05	official
30	seed-oss-36b-instruct	36.52	official
31	sensechat-turbo	33.43	official
32	ernie-5.0-preview	33.36	official
33	deepseek-r1	32.41	official
34	deepseek-v3	29.99	official
35	llama-4-maverick	29.74	official
36	hunyuan-t1	28.9	official
37	mistral-large-2	26.5	official
38	llama-3.3-70b-instruct	26.17	official
39	ernie-4.5-21b-a3b	24.31	official
40	gpt-4.1-mini	22.2	official
41	kimi-k2	18.45	official
42	pangu-pro-moe-72b-a16b	16.5	official
43	sensenova-v6.5-turbo	13.96	official

legality

Measures whether the model avoids illegal or public-safety harms across continuously updated safety prompts.

Rank	Model	Value	Provenance
1	claude-haiku-4.5	98.62	official
2	claude-sonnet-4.6	97.77	official
3	doubao-seed-1.6	93.2	official
4	gpt-5.2	89.24	official
5	gpt-oss-120b	86.06	official
6	deepseek-r1	85.68	official
7	qwen3.5-plus	84.85	official
8	intern-s1-pro	80.56	official
9	spark-x2	79.29	official
10	doubao-seed-2.0-pro	78.59	official
11	qwen3-235b-a22b	73.51	official
12	minimax-m2	67.96	official
13	kimi-k2.5	67.47	official
14	llama-3.3-70b-instruct	66.51	official
15	deepseek-v3.2	65.16	official
16	qwen3.5-397b-a17b	62.2	official
17	gemini-2.5-flash	56.07	official
18	mimo-v2-flash	52.92	official
19	gemini-3.1-pro-preview	50.36	official
20	sensechat-turbo	48.96	official
21	minimax-m2.5	48.6	official
22	longcat-flash-chat	46.38	official
23	intern-s1	42.89	official
24	mistral-large-2	42.83	official
25	glm-5	39.78	official
26	glm-4.6	39.39	official
27	kimi-k2	31.19	official
28	ernie-5.0-preview	30.82	official
29	ling-2.5-1t	30.6	official
30	llama-4-maverick	27.05	official
31	step-3.5-flash	26.76	official
32	hunyuan-t1	23.02	official
33	grok-3-mini	20.79	official
33	mistral-large-3	20.79	official
35	gpt-4.1-mini	19.69	official
36	seed-oss-36b-instruct	19.65	official
37	step-3	19.37	official
38	grok-4.1-fast	18.02	official
39	ernie-4.5-21b-a3b	15.07	official
40	pangu-pro-moe-72b-a16b	12.05	official
41	sensenova-v6.5-turbo	11.17	official
42	deepseek-v3	7.76	official
43	hunyuan-a13b-instruct	5.3	official

privacy

Measures whether the model avoids privacy violations or disclosure of personal information across continuously updated safety prompts.

Rank	Model	Value	Provenance
1	claude-haiku-4.5	97.17	official
2	qwen3.5-397b-a17b	94.2	official
3	claude-sonnet-4.6	89.26	official
4	gpt-5.2	87.9	official
5	spark-x2	86.96	official
6	qwen3.5-plus	79.84	official
7	doubao-seed-1.6	75.73	official
8	kimi-k2	73.81	official
9	longcat-flash-chat	73.3	official
10	qwen3-235b-a22b	72.57	official
11	sensechat-turbo	68.09	official
12	gemini-3.1-pro-preview	67.16	official
13	gpt-oss-120b	65.7	official
14	minimax-m2	64.5	official
15	intern-s1-pro	63.69	official
16	doubao-seed-2.0-pro	62.37	official
17	kimi-k2.5	60.89	official
18	glm-5	60.57	official
19	minimax-m2.5	59.81	official
20	deepseek-r1	59.73	official
21	ernie-5.0-preview	54.35	official
22	mimo-v2-flash	53.69	official
23	step-3.5-flash	46.86	official
24	ling-2.5-1t	44.71	official
25	grok-4.1-fast	39.48	official
26	deepseek-v3.2	37.7	official
27	sensenova-v6.5-turbo	35.54	official
28	hunyuan-t1	31.95	official
29	gemini-2.5-flash	31.6	official
30	step-3	30.7	official
31	llama-3.3-70b-instruct	30.35	official
32	grok-3-mini	30.02	official
33	glm-4.6	29.92	official
34	mistral-large-2	26.82	official
35	llama-4-maverick	26.17	official
36	seed-oss-36b-instruct	21.14	official
37	ernie-4.5-21b-a3b	19.99	official
38	intern-s1	19.05	official
39	gpt-4.1-mini	18.39	official
40	hunyuan-a13b-instruct	16.52	official
41	mistral-large-3	10.14	official
42	deepseek-v3	6.72	official
43	pangu-pro-moe-72b-a16b	5.84	official

psychological_health

Measures whether the model avoids responses that could harm a user’s psychological health across continuously updated safety prompts.

Rank	Model	Value	Provenance
1	gpt-5.2	92.91	official
2	claude-haiku-4.5	90.02	official
3	qwen3.5-397b-a17b	89.45	official
4	kimi-k2.5	85.12	official
5	qwen3.5-plus	80.56	official
6	gemini-3.1-pro-preview	76.44	official
7	glm-5	76.17	official
8	claude-sonnet-4.6	72.01	official
9	gpt-oss-120b	71.2	official
10	minimax-m2	69.47	official
11	qwen3-235b-a22b	69.27	official
12	step-3.5-flash	68.03	official
13	spark-x2	66.62	official
14	deepseek-v3.2	61.46	official
15	glm-4.6	60.3	official
16	deepseek-r1	58.87	official
17	doubao-seed-1.6	58.17	official
18	step-3	55.88	official
19	gemini-2.5-flash	54.27	official
20	longcat-flash-chat	52.22	official
21	doubao-seed-2.0-pro	51.42	official
22	minimax-m2.5	50.99	official
23	intern-s1-pro	50	official
24	ling-2.5-1t	47.77	official
25	ernie-5.0-preview	47.6	official
26	mimo-v2-flash	45.77	official
27	grok-4.1-fast	44.84	official
28	mistral-large-3	43.44	official
29	llama-4-maverick	35.37	official
30	kimi-k2	34.35	official
31	grok-3-mini	33.15	official
32	gpt-4.1-mini	30.95	official
33	pangu-pro-moe-72b-a16b	27.71	official
34	seed-oss-36b-instruct	27.32	official
35	ernie-4.5-21b-a3b	25.39	official
36	deepseek-v3	25.03	official
37	mistral-large-2	24.54	official
38	hunyuan-t1	24.18	official
39	intern-s1	19.6	official
40	llama-3.3-70b-instruct	17.08	official
41	sensechat-turbo	12.27	official
42	hunyuan-a13b-instruct	10.79	official
43	sensenova-v6.5-turbo	7.71	official