Mental & Emotional Safety ↗

SIM-VAIL

Whether a model responds safely to simulated vulnerable users across adaptive multi-turn conversations.

1. 100AnthropicClaude Sonnet 5
2. 100AnthropicClaude Opus 5.5
3. 100Zhipu AIGLM 5.3 Flash
4. 97xAIGrok 4.6
5. 97xAIGrok 4.5
6. 96OpenAIGPT-6 Astra
7. 96IInklingInkling
8. 95OpenAIGPT-5.6 Terra
9. 95GoogleGemini 3.6 Flash
10. 95PerplexityPerplexity Agent

Spiral-Bench

Whether a model pushes back on, or reinforces, delusional, manic or conspiratorial spirals over long chats.

1. 76OpenAIGPT-6 Astra
2. 72PerplexityPerplexity Agent
3. 72OpenAIGPT-5.6 Luna
4. 71OpenAIGPT-5.6 Terra
5. 70xAIGrok 4.6
6. 70AnthropicClaude Opus 5.5
7. 70Zhipu AIGLM 5.3 Flash
8. 69OpenAIGPT-6 Sol
9. 68AnthropicClaude Sonnet 5
10. 66IInklingInkling

Youth Safety ↗

KORA

Whether a model keeps children aged 7–17 safe across 25 risks in short conversations with a simulated child.

1. 82OpenAIGPT-6 Astra
2. 80OpenAIGPT-5.6 Terra
3. 78OpenAIGPT-5.6 Luna
4. 78OpenAIGPT-6 Sol
5. 75PerplexityPerplexity Agent
6. 73xAIGrok 4.6
7. 70AnthropicClaude Opus 5.5
8. 64IInklingInkling
9. 64Zhipu AIGLM 5.3 Flash
10. 63xAIGrok 4.5

Medical Advice Safety ↗

PatientSafetyBench

How harmful a model's answers are to risky medical questions asked from a patient's perspective.

1. 100xAIGrok 4.6
2. 99xAIGrok 4.5
3. 96OpenAIGPT-6 Sol
4. 96OpenAIGPT-6 Astra
5. 96PerplexityPerplexity Agent
6. 95DeepSeekDeepSeek V4 Flash
7. 95OpenAIGPT-5.6 Terra
8. 95IInklingInkling
9. 95OpenAIGPT-5.6 Luna
10. 94GoogleGemini 3.6 Flash

HealthBench-Hard

How well a model's reply in hard, realistic health conversations meets physician-written rubric criteria.

1. 39AnthropicClaude Opus 5.5
2. 34OpenAIGPT-5.6 Terra
3. 33OpenAIGPT-6 Astra
4. 33IInklingInkling
5. 31xAIGrok 4.5
6. 31OpenAIGPT-5.6 Luna
7. 30PerplexityPerplexity Agent
8. 28GoogleGemini 3.6 Flash
9. 24xAIGrok 4.6
10. 23Zhipu AIGLM 5.3 Flash

Manipulation ↗

ELEPHANT

How often a model is socially sycophantic in advice: excessive validation, indirectness and accepting the user's framing.

1. 47IInklingInkling
2. 47GoogleGemini 3.6 Flash
3. 46PerplexityPerplexity Agent
4. 44xAIGrok 4.6
5. 41AnthropicClaude Sonnet 5
6. 38xAIGrok 4.5
7. 37DeepSeekDeepSeek V4 Flash
8. 36Zhipu AIGLM 5.3 Flash
9. 34OpenAIGPT-5.6 Luna
10. 33OpenAIGPT-5.6 Terra

DarkBench

How often a chatbot's reply shows one of six manipulative dark patterns, such as brand bias or user retention.

1. 65AnthropicClaude Sonnet 5
2. 57IInklingInkling
3. 57xAIGrok 4.6
4. 53Zhipu AIGLM 5.3 Flash
5. 52OpenAIGPT-5.6 Terra
6. 52OpenAIGPT-6 Sol
7. 51xAIGrok 4.5
8. 50OpenAIGPT-6 Astra
9. 49OpenAIGPT-5.6 Luna
10. 48PerplexityPerplexity Agent

HumanAgencyBench (Autonomy)

Whether a model leaves a person's values and important decisions to them rather than steering them (HumanAgencyBench's Avoid Value Manipulation and Defer Important Decisions dimensions).

1. 58xAIGrok 4.5
2. 57xAIGrok 4.6
3. 55OpenAIGPT-6 Astra
4. 52OpenAIGPT-6 Sol
5. 50AnthropicClaude Sonnet 5
6. 45AnthropicClaude Opus 5.5
7. 44GoogleGemini 3.6 Flash
8. 41IInklingInkling
9. 37OpenAIGPT-5.6 Terra
10. 31Zhipu AIGLM 5.3 Flash

Security ↗

ASK — AI Scam Knowledge

Whether a model spots a scam the user describes and warns them, as scam cues increase.

1. 72AnthropicClaude Sonnet 5
2. 68GoogleGemini 3.6 Flash
3. 67IInklingInkling
4. 67Zhipu AIGLM 5.3 Flash
5. 62PerplexityPerplexity Agent
6. 58OpenAIGPT-5.6 Terra
7. 58OpenAIGPT-6 Sol
8. 58AnthropicClaude Opus 5.5
9. 58xAIGrok 4.6
10. 56DeepSeekDeepSeek V4 Flash

Bias & Fairness ↗

FairMT-Bench

Whether a model produces biased content by the fifth turn of dialogues designed to draw out social bias.

1. 90OpenAIGPT-6 Sol
2. 85OpenAIGPT-6 Astra
3. 79xAIGrok 4.6
4. 78OpenAIGPT-5.6 Terra
5. 76OpenAIGPT-5.6 Luna
6. 74AnthropicClaude Opus 5.5
7. 73PerplexityPerplexity Agent
8. 69xAIGrok 4.5
9. 60Zhipu AIGLM 5.3 Flash
10. 58AnthropicClaude Sonnet 5

Privacy / Confidentiality ↗

ConfAIde

Whether a model reveals private information where social norms say it should stay confidential.

1. 99AnthropicClaude Opus 5.5
2. 99Zhipu AIGLM 5.3 Flash
3. 98OpenAIGPT-6 Sol
4. 98OpenAIGPT-6 Astra
5. 98xAIGrok 4.5
6. 98OpenAIGPT-5.6 Terra
7. 98AnthropicClaude Sonnet 5
8. 98xAIGrok 4.6
9. 98PerplexityPerplexity Agent
10. 98OpenAIGPT-5.6 Luna

Misinformation ↗

SimpleQA Verified

How accurately a model answers short factual questions from its own knowledge, and whether it abstains instead of guessing.

1. 98PerplexityPerplexity Agent
2. 75OpenAIGPT-6 Astra
3. 74AnthropicClaude Opus 5.5
4. 69GoogleGemini 3.6 Flash
5. 64OpenAIGPT-6 Sol
6. 51xAIGrok 4.5
7. 51xAIGrok 4.6
8. 43OpenAIGPT-5.6 Terra
9. 42IInklingInkling
10. 40OpenAIGPT-5.6 Luna

SYCON-Bench

How many turns a model holds its position under repeated user pushback before giving in.

1. 99xAIGrok 4.5
2. 99xAIGrok 4.6
3. 99DeepSeekDeepSeek V4 Flash
4. 99GoogleGemini 3.6 Flash
5. 99IInklingInkling
6. 97Zhipu AIGLM 5.3 Flash
7. 89OpenAIGPT-6 Astra
8. 89OpenAIGPT-5.6 Luna
9. 89OpenAIGPT-5.6 Terra
10. 88AnthropicClaude Opus 5.5

HumanAgencyBench (Correct Misinformation)

Whether a model notices and corrects a false claim built into a user's request (HumanAgencyBench's Correct Misinformation dimension).

1. 96AnthropicClaude Opus 5.5
2. 90OpenAIGPT-6 Astra
3. 89PerplexityPerplexity Agent
4. 85OpenAIGPT-6 Sol
5. 82OpenAIGPT-5.6 Terra
6. 82Zhipu AIGLM 5.3 Flash
7. 79AnthropicClaude Sonnet 5
8. 72xAIGrok 4.6
9. 68xAIGrok 4.5
10. 63OpenAIGPT-5.6 Luna

Rule Following ↗

SystemCheck / RealGuardrails

Whether a model obeys the guardrails in a real-world system prompt when users push against them.

1. 94GoogleGemini 3.6 Flash
2. 89IInklingInkling
3. 86xAIGrok 4.5
4. 85OpenAIGPT-6 Sol
5. 82AnthropicClaude Opus 5.5
6. 82OpenAIGPT-5.6 Terra
7. 80xAIGrok 4.6
8. 80OpenAIGPT-6 Astra
9. 76OpenAIGPT-5.6 Luna
10. 76Mistral AIMistral Medium 3.5

AgentIF

Whether a model satisfies every constraint in long, realistic instructions from agent applications.

1. 63AnthropicClaude Opus 5.5
2. 63Zhipu AIGLM 5.3 Flash
3. 62GoogleGemini 3.6 Flash
4. 61DeepSeekDeepSeek V4 Flash
5. 60AnthropicClaude Sonnet 5
6. 60OpenAIGPT-6 Astra
7. 59xAIGrok 4.6
8. 58OpenAIGPT-5.6 Terra
9. 58OpenAIGPT-5.6 Luna
10. 58xAIGrok 4.5