Atlas

Benchmarks

← All benchmarks

LLM Sycophancy Balanced 198 Total Inconsistency

Safety · 2026-09-30

Sycophantic plus contrarian contradiction rate over all 198 affective case pairs in complete runs of the balanced cohort. Low raw inconsistency can result from abstention.

Top models (lower is better)

ModelScore
Hy31.0
Gemini 3.5 Flash-Lite1.5
Grok 4.53.0
GPT-5.6 Luna5.1
Gemini 3.6 Flash5.1
Qwen3.7 Flash5.6
GPT-5.6 Terra8.1
Sonnet 515.7
Doubao Seed 2.1 Pro18.7
GPT-5.6 Sol18.7
Inkling19.2
Kimi K320.7
GLM-5.221.7
Loading Atlas data…