Atlas

Benchmarks

← All benchmarks

Phare Harm Resistance

Safety · 2025-05-16

Phare module score for resistance to harmful guidance. Higher is better.

Top models (higher is better)

ModelScore
Opus 4.6100.0
Haiku 4.599.9
Sonnet 4.699.9
Sonnet 599.6
Sonnet 4.599.0
GPT-5.598.9
GPT-5 Mini98.3
Opus 4.598.3
GPT-5 Nano97.4
Kimi K2.597.2
Kimi K2.697.0
GPT-597.0
Gemini 3.1 Flash-Lite96.9
GPT-5.196.9
GPT-5.296.9
Gemini 1.5 Pro96.8
Gemma 4 31B IT96.5
Opus 4.196.3
Gemini 3.5 Flash95.9
Qwen3.7 Max Preview95.6
Claude 3.7 Sonnet95.5
DeepSeek-V4-Pro95.4
Claude 3.5 Sonnet (Oct 2024)95.4
Qwen3 Max (rolling alias)95.4
Haiku 3.595.4
Gemini 3.1 Pro Preview95.2
DeepSeek-R1-052895.2
DeepSeek-V4-Flash95.1
DeepSeek-V3.194.4
Qwen3.7-Plus94.4
Gemini 2.0 Flash94.3
Qwen-Plus (2025-01-25)94.1
gpt-oss-120b93.8
GLM-5.293.7
Gemini 2.5 Flash93.7
Grok 4.393.5
Gemini 3 Pro Preview93.5
DeepSeek-V3-032492.8
GPT-4o92.7
Gemma 3 12B IT92.7
Loading Atlas data…