TukaBench - Refusal Rate: leaderboard
Refusal rate on the TukaBench (2026) harmful prompts in six African languages (Amharic, Hausa, Igbo, Chichewa, Kiswahili and Yoruba): the share of responses a GPT-4.1 judge labels explicit refusals rather than jailbroken or deflected (off-target) answers, averaged over the three prompt sets. Scores are percentages; higher is better.
Metric: Refusal rate (%; share of harmful prompts explicitly refused, 100 minus the attack success and deflection rates) averaged over six African languages (Amharic, Hausa, Igbo, Chichewa, Kiswahili, Yoruba; isiXhosa excluded), TukaBench jailbreak prompts (JailbreakBench-derived Afri-JBB-Harm and Afri-JBB-Culture plus African-authored AfriJail-Mono), direct prompting, greedy decoding, GPT-4.1 judge labelling each response Jailbroken, Refused or Deflected; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 83.1 | #176 |
| 2 | Claude Opus 4.8 | 79.9 | #56 |
| 3 | Grok 4 | 78.7 | #307 |
| 4 | Grok 4.3 | 78.1 | #333 |
| 5 | Qwen 3.5 27B | 77.3 | #359 |
| 6 | Gemma 3 27B | 75.5 | #1631 |
| 7 | Gemma 4 31B | 68.8 | #394 |
| 8 | GPT-OSS-120B | 68.4 | #700 |
| 9 | Claude Sonnet 5 | 63.4 | #111 |
| 10 | DeepSeek V3.2 | 61.1 | #406 |
| 11 | GPT-4o | 60.7 | #741 |
| 12 | Llama 4 Maverick | 60.1 | #1099 |
| 13 | Gemini 3.1 Pro (Preview) | 53.6 | #94 |
| 14 | Grok 3 | 45.8 | #602 |
| 15 | Claude Haiku 4.5 | 43.4 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tukabench-refusal-rate · How It Works · Data refreshed daily, snapshot 2026-11-13.