Atlas

Benchmarks

← All benchmarks

PCT Political Inconsistency (CAIS)

Safety · 2026-05-21

Polarized Contrastive Pairs evaluates covert political manipulation by comparing model responses to matched prompts about opposing political subjects. SAFE reports Political Inconsistency, the complement of Political Consistency across 1,000 paired evaluations, so lower scores indicate more symmetric sentiment and helpfulness.

Top models (lower is better)

ModelScore
Muse Spark 1.127.9
Grok 4 Fast30.8
Grok 4.1 Fast32.5
GPT-535.3
Gemini 2.5 Pro37.3
GPT-5.539.0
o339.0
Claude Fable 539.8
Grok 4.2039.9
GPT-5.6 Sol41.7
Gemini 3.1 Pro Preview43.3
Kimi K2.643.6
DeepSeek-V4-Pro44.0
GPT-5.6 Luna44.7
GPT-5.6 Terra46.2
GPT-5.147.9
Opus 4.748.2
Opus 4.549.4
Grok 449.5
GLM-5.250.0
Grok 4.550.2
Grok 4.351.6
Opus 4.651.7
Gemini 3.1 Flash-Lite Preview51.9
Kimi K352.4
Gemini 3.5 Flash53.4
Opus 4.853.8
GPT-5.257.0
GPT-5.459.4
GPT-4o (2024-11-20)59.6
GPT-5.4 Mini61.1
Haiku 4.562.2
GPT-5.4 Nano63.3
Loading Atlas data…