Bullshit Benchmark: leaderboard
Tests whether models detect nonsense. Questions use fake precision, confident nonsense framing, cross-domain concept stitching, and other deceptive techniques. Rated green/amber/red by multiple judges.
Metric: BS Detection Rate (%). Source: petergpt.github.io. Status: years away from saturation. 188 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.8 | 96.4 |
| 2 | Claude Sonnet 4.6 | 94.5 |
| 3 | Claude Opus 4.6 | 92.7 |
| 4 | Claude Opus 4.5 | 90.9 |
| 5 | Claude Haiku 4.5 | 87.3 |
| 6 | Qwen 3.8 Max | 87.3 |
| 7 | Claude Opus 4.7 | 80 |
| 8 | Qwen 3.8 27B | 76.4 |
| 9 | Grok 4.20 Multi-Agent | 67.3 |
| 10 | Claude Sonnet 4.5 | 65.5 |
| 11 | Qwen 3.5 397B A17B | 65.5 |
| 12 | Claude Sonnet 5 | 65.5 |
| 13 | Nemotron 3 Super | 65.5 |
| 14 | Grok 4.6 | 63.6 |
| 15 | Grok 4.20 | 61.8 |
Interactive version: theaggregate.ai/benchmark?slug=bullshit-benchmark · How It Works · Data refreshed daily, snapshot 2026-09-05.