Evaluation profile

PHARE

4sub-evals

1.09%total index weight

3components

Within-component eval weight: Fairness 5.37% · Truthfulness 2.5% · Misuse resistance 1.76%.

Model score (higher is better)Predicted score

About this eval

Broad safety across hallucination, harmfulness, out-of-scope handling, and bias.

Included in the behavior ranking.

Sub-evals

Measure	Component	Direction	Total index weight	Within-component weight
bias_resistance_diagnosticphare/phare.csv:bias_resistance_diagnosticMeasures whether the model avoids reproducing stereotypes or biased treatment across the benchmark’s multilingual bias tasks.	fairness_nondiscrimination:1.000phare	Higher is better	0.537%	Fairness 5.37%
hallucination_resistance_diagnosticphare/phare.csv:hallucination_resistance_diagnosticMeasures whether the model remains factually accurate, rejects misinformation, debunks false claims, and handles incomplete tool inputs without fabrication.	truthfulness_honesty:1.000phare	Higher is better	0.375%	Truthfulness 2.5%
harm_resistance_diagnosticphare/phare.csv:harm_resistance_diagnosticMeasures whether the model avoids generating responses that could cause harm across the benchmark’s multilingual harmfulness tasks.	ordinary_harm_misuse_resistance:1.000phare	Higher is better	0.0892%	Misuse resistance 0.892%
jailbreak_resistance_diagnosticphare/phare.csv:jailbreak_resistance_diagnosticMeasures whether the model maintains its safety behavior when users apply jailbreak and prompt-injection techniques.	ordinary_harm_misuse_resistance:1.000phare	Higher is better	0.0873%	Misuse resistance 0.873%

bias_resistance_diagnostic

Measures whether the model avoids reproducing stereotypes or biased treatment across the benchmark’s multilingual bias tasks.

Rank	Model	Value	Provenance
1	gpt-4.1-mini	0.8812	official
2	grok-4-fast	0.8026	official
3	llama-3.1-405b-instruct	0.7523	official
4	mistral-small-3.2	0.739	official
5	llama-4-maverick	0.7365	official
6	claude-haiku-4.5	0.7066	official
7	llama-4-scout	0.671	official
8	deepseek-v3.1	0.6517	official
9	claude-opus-4.5	0.632	official
10	mistral-large-3	0.6272	official
11	mistral-medium-3.5	0.6098	official
12	qwen3-8b	0.5864	official
13	deepseek-v3	0.5751	official
14	qwen-plus	0.557	official
15	qwen3.7-plus	0.5496	official
16	gemini-3-pro-preview	0.5365	official
17	qwen3-30b-a3b-instruct	0.5357	official
18	gemini-2.0-flash	0.5351	official
19	kimi-k2.6	0.5307	official
20	gpt-4.1	0.5241	official
21	qwen3.7-max	0.524	official
22	gemini-2.5-flash	0.5174	official
23	glm-5.2	0.5127	official
24	gpt-4o	0.5092	official
25	magistral-medium	0.5075	official
26	claude-sonnet-4.5	0.4914	official
27	gemini-3.1-pro	0.481	official
28	magistral-small	0.4802	official
29	grok-4.3	0.47	official
30	gpt-5.1	0.4677	official
31	gpt-5-mini	0.4641	official
32	grok-3-mini	0.464	official
33	command-a	0.4559	official
34	gemini-2.5-flash-lite	0.4553	official
35	llama-3.3-70b-instruct	0.4539	official
36	qwen3-max	0.4477	official
37	gemini-3.1-flash-lite	0.4457	official
38	llama-3.1-8b-instruct	0.4415	official
39	claude-opus-4.1	0.4361	official
40	qwen2.5-max	0.4295	official
41	gemini-2.0-flash-lite	0.4165	official
42	claude-sonnet-4.6	0.4128	official
43	claude-sonnet-5	0.3988	official
44	mistral-large-2	0.397	official
45	mistral-medium-3.1	0.3966	official
46	gpt-oss-120b	0.3884	official
47	gpt-5.2	0.3851	official
48	gpt-5.5	0.3829	official
49	claude-3.5-haiku	0.3808	official
50	gemma-3-27b-it	0.3801	official
51	gpt-4o-mini	0.379	official
52	gemini-2.5-pro	0.3652	official
53	gpt-4.1-nano	0.3622	official
54	gemma-4-31b-it	0.3522	official
55	gpt-5-nano	0.347	official
56	claude-3.7-sonnet	0.3377	official
57	claude-opus-4.6	0.3351	official
58	grok-2	0.331	official
59	deepseek-v4-flash	0.323	official
60	gemma-3-12b	0.3214	official
61	grok-4	0.3116	official
62	kimi-k2.5	0.29	official
63	gpt-5	0.2856	official
64	deepseek-r1	0.2549	official
65	grok-3	0.2324	official
66	deepseek-v4-pro	0.1714	official

hallucination_resistance_diagnostic

Measures whether the model remains factually accurate, rejects misinformation, debunks false claims, and handles incomplete tool inputs without fabrication.

Rank	Model	Value	Provenance
1	claude-opus-4.5	0.8823	official
2	claude-opus-4.6	0.8776	official
3	gemini-3.1-pro	0.8759	official
4	claude-sonnet-4.5	0.87	official
5	claude-3.5-sonnet	0.8671	official
6	claude-opus-4.1	0.8619	official
7	claude-sonnet-5	0.861	official
8	claude-sonnet-4.6	0.8551	official
9	claude-3.7-sonnet	0.8517	official
10	claude-haiku-4.5	0.8356	official
11	kimi-k2.6	0.8278	official
12	gemini-3.5-flash	0.8218	official
13	deepseek-v4-flash	0.8188	official
14	gpt-5.1	0.8184	official
15	qwen3.7-max	0.8114	official
16	gemini-3-pro-preview	0.8102	official
17	grok-4	0.8073	official
18	kimi-k2.5	0.8069	official
19	gemini-3.1-flash-lite	0.8057	official
20	gpt-5.5	0.7978	official
21	deepseek-v4-pro	0.7964	official
22	gpt-5-mini	0.7958	official
23	grok-4.3	0.7946	official
24	gpt-4o	0.7852	official
25	glm-5.2	0.7823	official
26	gemini-1.5-pro	0.7807	official
27	claude-3.5-haiku	0.7804	official
28	gpt-4.1	0.7764	official
29	gpt-5.2	0.7713	official
30	gemma-4-31b-it	0.7662	official
31	gpt-5-nano	0.7637	official
32	qwen3.7-plus	0.7626	official
33	mistral-large-2	0.7562	official
34	qwen-plus	0.7558	official
35	qwen3-max	0.752	official
36	grok-3	0.7512	official
37	gemini-2.5-pro	0.7482	official
38	gpt-5	0.7458	official
39	deepseek-r1	0.7289	official
40	gemini-2.5-flash	0.7213	official
41	command-a	0.7189	official
42	gemini-2.0-flash	0.7159	official
43	llama-4-maverick	0.7147	official
44	mistral-small-3.1	0.7	official
45	qwen2.5-max	0.698	official
46	mistral-medium-3.1	0.6963	official
47	llama-3.3-70b-instruct	0.6883	official
48	grok-3-mini	0.6862	official
49	magistral-medium	0.6819	official
50	gemini-2.5-flash-lite	0.6809	official
51	mistral-large-3	0.6803	official
52	gpt-4.1-mini	0.6781	official
53	gpt-4.1-nano	0.6774	official
54	mistral-medium-3.5	0.677	official
55	llama-3.1-405b-instruct	0.674	official
56	grok-2	0.668	official
57	deepseek-v3	0.6675	official
58	gpt-oss-120b	0.6608	official
59	gpt-4o-mini	0.6597	official
60	llama-3.1-8b-instruct	0.6381	official
61	mistral-small-3.2	0.634	official
62	grok-4-fast	0.634	official
63	gemini-2.0-flash-lite	0.6321	official
64	magistral-small	0.6244	official
65	deepseek-v3.1	0.6158	official
66	qwen3-8b	0.6038	official
67	gemma-3-27b-it	0.603	official
68	qwen3-30b-a3b-instruct	0.5988	official
69	llama-4-scout	0.582	official
70	gemma-3-12b	0.5386	official

harm_resistance_diagnostic

Measures whether the model avoids generating responses that could cause harm across the benchmark’s multilingual harmfulness tasks.

Rank	Model	Value	Provenance
1	claude-opus-4.6	1	official
2	claude-sonnet-4.6	0.9993	official
3	claude-haiku-4.5	0.9993	official
4	claude-sonnet-5	0.9959	official
5	claude-sonnet-4.5	0.9905	official
6	gpt-5.5	0.9892	official
7	gpt-5-mini	0.9829	official
8	claude-opus-4.5	0.9825	official
9	gpt-5-nano	0.9741	official
10	kimi-k2.5	0.972	official
11	kimi-k2.6	0.97	official
12	gpt-5	0.9697	official
13	gemini-3.1-flash-lite	0.9694	official
14	gpt-5.1	0.9692	official
15	gpt-5.2	0.9688	official
16	gemini-1.5-pro	0.9684	official
17	gemma-4-31b-it	0.9651	official
18	claude-opus-4.1	0.9631	official
19	gemini-3.5-flash	0.9594	official
20	qwen3.7-max	0.9558	official
21	claude-3.7-sonnet	0.9552	official
22	deepseek-v4-pro	0.9541	official
23	qwen3-max	0.954	official
24	claude-3.5-sonnet	0.954	official
25	claude-3.5-haiku	0.9536	official
26	gemini-3.1-pro	0.9519	official
27	deepseek-r1	0.9515	official
28	deepseek-v4-flash	0.9507	official
29	deepseek-v3.1	0.9443	official
30	qwen3.7-plus	0.9443	official
31	gemini-2.0-flash	0.943	official
32	qwen-plus	0.9414	official
33	gpt-oss-120b	0.9375	official
34	glm-5.2	0.9373	official
35	gemini-2.5-flash	0.9366	official
36	grok-4.3	0.9355	official
37	gemini-3-pro-preview	0.935	official
38	gpt-4o	0.9266	official
39	gemma-3-12b	0.9265	official
40	mistral-medium-3.1	0.9232	official
41	gpt-4.1	0.923	official
42	gemini-2.5-pro	0.9218	official
43	grok-2	0.9144	official
44	command-a	0.9136	official
45	gemma-3-27b-it	0.9136	official
46	mistral-medium-3.5	0.9115	official
47	mistral-small-3.1	0.9091	official
48	deepseek-v3	0.909	official
49	grok-3-mini	0.9047	official
50	qwen2.5-max	0.8989	official
51	grok-3	0.8968	official
52	mistral-large-2	0.8938	official
53	llama-4-maverick	0.8925	official
54	mistral-large-3	0.8806	official
55	mistral-small-3.2	0.8787	official
56	qwen3-8b	0.8737	official
57	llama-3.1-405b-instruct	0.8649	official
58	llama-3.3-70b-instruct	0.8604	official
59	gemini-2.0-flash-lite	0.8514	official
60	magistral-medium	0.8452	official
61	gpt-4.1-mini	0.8339	official
62	llama-3.1-8b-instruct	0.8306	official
63	qwen3-30b-a3b-instruct	0.8176	official
64	grok-4-fast	0.8134	official
65	llama-4-scout	0.8104	official
66	gemini-2.5-flash-lite	0.7915	official
67	gpt-4o-mini	0.7729	official
68	magistral-small	0.7623	official
69	gpt-4.1-nano	0.7254	official
70	grok-4	0.7177	official

jailbreak_resistance_diagnostic

Measures whether the model maintains its safety behavior when users apply jailbreak and prompt-injection techniques.

Rank	Model	Value	Provenance
1	claude-sonnet-5	0.8684	official
2	claude-opus-4.1	0.8135	official
3	claude-opus-4.5	0.7983	official
4	claude-haiku-4.5	0.7851	official
5	llama-3.1-405b-instruct	0.7655	official
6	claude-opus-4.6	0.7524	official
7	claude-sonnet-4.5	0.7523	official
8	qwen3.7-max	0.7316	official
9	gpt-5.2	0.7161	official
10	kimi-k2.6	0.714	official
11	qwen3.7-plus	0.7058	official
12	gpt-5-nano	0.6964	official
13	claude-sonnet-4.6	0.6939	official
14	gpt-5	0.6868	official
15	gpt-5-mini	0.686	official
16	gpt-oss-120b	0.659	official
17	gpt-5.1	0.6576	official
18	gpt-5.5	0.6513	official
19	gemini-3-pro-preview	0.6506	official
20	grok-4	0.6485	official
21	claude-3.5-haiku	0.6482	official
22	claude-3.7-sonnet	0.6345	official
23	kimi-k2.5	0.613	official
24	gemma-4-31b-it	0.6122	official
25	gpt-4o	0.6048	official
26	llama-3.1-8b-instruct	0.5881	official
27	glm-5.2	0.5646	official
28	grok-4.3	0.545	official
29	gpt-4.1-nano	0.533	official
30	gpt-4o-mini	0.5214	official
31	gemini-3.1-pro	0.5164	official
32	qwen-plus	0.4998	official
33	qwen3-8b	0.4913	official
34	llama-4-maverick	0.4899	official
35	llama-4-scout	0.4893	official
36	qwen3-max	0.4835	official
37	gemini-3.1-flash-lite	0.4772	official
38	llama-3.3-70b-instruct	0.4768	official
39	qwen2.5-max	0.4626	official
40	gpt-4.1	0.461	official
41	magistral-medium	0.4561	official
42	deepseek-v4-flash	0.4549	official
43	gemma-3-12b	0.4427	official
44	deepseek-v4-pro	0.4321	official
45	gemini-2.5-flash-lite	0.4284	official
46	gemini-3.5-flash	0.4264	official
47	gpt-4.1-mini	0.4255	official
48	qwen3-30b-a3b-instruct	0.4209	official
49	grok-4-fast	0.412	official
50	deepseek-r1	0.4067	official
51	gemma-3-27b-it	0.4007	official
52	gemini-2.0-flash-lite	0.3926	official
53	gemini-2.5-pro	0.3867	official
54	deepseek-v3.1	0.3817	official
55	mistral-medium-3.5	0.3746	official
56	gemini-2.0-flash	0.3735	official
57	magistral-small	0.3709	official
58	gemini-2.5-flash	0.3705	official
59	command-a	0.3482	official
60	grok-3-mini	0.3436	official
61	mistral-small-3.2	0.3289	official
62	deepseek-v3	0.3107	official
63	mistral-medium-3.1	0.3083	official
64	mistral-large-2	0.3058	official
65	grok-2	0.2697	official
66	grok-3	0.2577	official
67	mistral-large-3	0.249	official