Atlas

Benchmarks

← All benchmarks

BullshitBench v2 Leaderboard

Safety · 2026-02-24

Clear Pushback rate on nonsense prompts; higher scores mean the model rejects more bad claims instead of accepting them.

Top models (higher is better)

ModelScore
Opus 4.895.0
Sonnet 4.691.0
Opus 4.590.0
Opus 4.687.0
Opus 4.783.0
Sonnet 580.0
Sonnet 4.579.0
Qwen3.5 397B A17B78.0
Haiku 4.577.0
Claude Opus 573.0
Kimi K373.0
Qwen3.6 Plus (2026-04-02)72.0
Qwen3.7-Max71.0
Grok 4.2067.0
Kimi K2.665.0
Qwen3 Max Thinking (2026-01-23)63.0
MiniMax M363.0
MiMo-V2.5-Pro62.0
Claude Fable 554.0
Grok 4.554.0
Nemotron 3 Super 120B A12B54.0
GPT-5.6 Terra53.0
Kimi K2.552.0
Grok 4.350.0
Haiku 3.550.0
Claude 3.7 Sonnet49.0
Nemotron 3 Ultra 550B A55B49.0
GPT-5.448.0
Gemini 3 Pro Preview48.0
GPT-5.6 Sol47.0
GPT-5.547.0
GPT-5.2-Codex45.0
Claude 3.5 Sonnet (Oct 2024)45.0
GPT-5.1 Instant45.0
Opus 4.143.0
GPT-5.6 Luna40.0
GPT-5.3 Instant40.0
GPT-5-Codex39.0
GPT-5.238.0
Gemini 3.1 Pro Preview37.0
Loading Atlas data…