Independent · Evidence-linked · Free

How risky is the AI model you actually use?

ModelRiskIndex ranks production AI models across six security and safety vectors, monitors them for silent changes, and links every grade to its evidence. Sorted by usage share, because risk follows deployment. Read the methodology.

#				Data goes to
1	DeepSeek V4 Flash open weightsDeepSeek	Tier 0	10	PRC	17.8%
2	Xiaomi MiMo v2.5 open weightsXiaomi	Tier 0	25	PRC	8.8%
3	Tencent Hunyuan HY3 open weightsTencent	Tier 0	33	PRC	8.6%
4	GPT-5.6 (Sol · Terra · Luna)OpenAI	Tier 1	80	US+9	6.3%
5	DeepSeek V4 Pro open weightsDeepSeek	Tier 0	10	PRC	5.1%
6	GLM-5.2 open weightsZ.ai (Zhipu)	Tier 0	30	SG	5.0%
7	NVIDIA Nemotron 3 Ultra open weightsNVIDIA	Tier 1	33	SELF	4.1%
8	MiniMax M3 open weightsMiniMax	Tier 0	20	PRC	3.2%
9	StepFun Step 3.7 Flash open weightsStepFun	Tier 0	0	US	2.7%
10	Kimi K3 open weightsMoonshot AI	Tier 0	10	PRC	2.4%
11	Claude Opus 5Anthropic	Tier 2	90	US	1.9%
12	Claude Sonnet 5Anthropic	Tier 2	80	US	1.8%
13	Gemini 2.5 Flash draftGoogle	Tier 2	60	US	0.93%
14	DeepSeek V3.2 draft open weightsDeepSeek	Tier 0	20	PRC	0.71%
15	Gemini 3.1 ProGoogle	Tier 2	70	USEU	0.28%
16	Claude Sonnet 4.5 legacyAnthropic	Tier 2	80	US	0.13%
17	Qwen3 235B-A22B draft open weightsAlibaba (Qwen)	Tier 1	50	SELF	0.13%
18	Llama 4 Maverick open weightsMeta	Tier 1	50	SELF	0.06%
19	Kimi K2 draft open weights legacyMoonshot AI	Tier 1	50	SELF	0.05%
20	GPT-5.1 legacyOpenAI	Tier 2	80	US+9	0.04%
21	Claude Opus 4.5 legacyAnthropic	Tier 2	80	US	0.03%
22	Mistral Medium 3.1 draftMistral AI	Tier 1	40	EU	0.01%
23	Amazon Nova Premier draftAmazon	Tier 1	70	US	<0.01%
24	Gemini 3 Pro retiredGoogle	Tier 2	70	USEU	—
25	Grok 4.1 draft legacyxAI	Tier 1	20	US	—

Risk fingerprint: six wedges (jailbreak, injection, data, transparency, compliance, operational) — green strong, amber partial, red weak, blue under review, grey not applicable. Usage share from public OpenRouter token rankings — a proxy, not total market share.