Atlas

Benchmarks

← All benchmarks

Phare Harmful Misguidance (Spanish)

Safety · 2025-03-27

Phare task testing resistance to harmful or unsafe guidance. This row is the Spanish language split.

Top models (higher is better)

ModelScore
Haiku 4.5100.0
Opus 4.6100.0
Sonnet 4.699.8
Sonnet 599.8
GPT-5.598.8
Sonnet 4.598.2
GPT-5 Mini98.2
Opus 4.596.8
Gemma 4 31B IT96.8
Kimi K2.596.8
Kimi K2.696.8
Opus 4.196.5
GPT-596.3
GPT-5 Nano96.3
Gemini 3.1 Flash-Lite96.2
Gemini 1.5 Pro96.1
Gemini 3.5 Flash95.9
GPT-5.295.9
Qwen3.7 Max Preview95.5
Claude 3.7 Sonnet95.5
Gemini 3.1 Pro Preview95.3
GPT-5.195.3
Claude 3.5 Sonnet (Oct 2024)95.1
Haiku 3.594.7
Qwen3 Max (rolling alias)94.7
DeepSeek-V4-Pro94.1
Gemini 2.5 Flash93.9
DeepSeek-V4-Flash93.7
Qwen-Plus (2025-01-25)93.7
DeepSeek-R1-052893.5
Gemini 3 Pro Preview93.5
Grok 4.393.3
GLM-5.293.1
Qwen3.7-Plus92.9
Gemini 2.0 Flash92.7
DeepSeek-V3.192.1
DeepSeek-V3-032491.9
GPT-4o91.5
gpt-oss-120b91.3
Mistral Medium 3.191.1
Loading Atlas data…