ERR-EVAL - Trick Questions: leaderboard
The "Trick Questions" track of ERR-EVAL, which scores whether a model notices a defective prompt and says what it cannot answer instead of inventing it. This track feeds questions built on a false premise.
Metric: Track score. Source: err-eval.gustycube.com. Saturation forecast: Around April 2027. 60 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 6.8 | #83 |
| 2 | GPT-5 Mini | 6.4 | #193 |
| 3 | Claude Opus 4.5 | 6.2 | #87 |
| 4 | Grok 4.20 Multi-Agent | 6.2 | #134 |
| 5 | Claude Sonnet 4.5 | 6 | #148 |
| 6 | GPT-5.3 | 6 | #104 |
| 7 | GPT-5.3 Codex | 6 | #72 |
| 8 | GPT-5 | 5.8 | #105 |
| 9 | Kimi K2.5 | 5.8 | #152 |
| 10 | Grok 4 | 5.8 | #189 |
| 11 | DeepSeek R1 | 5.8 | #262 |
| 12 | DeepSeek R1 0528 | 5.8 | #232 |
| 13 | Grok 4.20 | 5.8 | #179 |
| 14 | MiniMax-M2.1 | 5.8 | #299 |
| 15 | Claude Opus 4.6 | 5.6 | #67 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=err-eval-trick-questions · How It Works · Data refreshed daily, snapshot 2026-11-05.