TukaBench - Benign Compliance: leaderboard
Over-refusal check of TukaBench (2026): the share of 100 benign JailbreakBench prompts, in English and six African languages, that a model answers rather than refuses or deflects, judged by GPT-4.1. Scores are percentages; higher is better.
Metric: Compliance rate (%; share of the 100 benign JailbreakBench prompts answered rather than refused or deflected) averaged over English and six African languages (isiXhosa excluded), direct prompting, greedy decoding, GPT-4.1 judge; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.3 | 40.7 | #333 |
| 2 | Grok 3 | 39.1 | #602 |
| 3 | Claude Haiku 4.5 | 39 | #568 |
| 4 | GPT-4o | 33 | #741 |
| 5 | Llama 4 Maverick | 31.9 | #1099 |
| 6 | Claude Opus 4.8 | 31.4 | #56 |
| 7 | DeepSeek V3.2 | 30.1 | #406 |
| 8 | GPT-5.2 | 22.9 | #176 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tukabench-benign-compliance · How It Works · Data refreshed daily, snapshot 2026-11-13.