TukaBench - Refusal Rate: leaderboard
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 December 2026. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.2 | 83.1 |
| 2 | Claude Opus 4.8 | 79.9 |
| 3 | Grok 4 | 78.7 |
| 4 | Grok 4.3 | 78.1 |
| 5 | Qwen 3.5 27B | 77.3 |
| 6 | Gemma 3 27B | 75.5 |
| 7 | Gemma 4 31B | 68.8 |
| 8 | GPT-OSS-120B | 68.4 |
| 9 | Claude Sonnet 5 | 63.4 |
| 10 | DeepSeek V3.2 | 61.1 |
| 11 | GPT-4o | 60.7 |
| 12 | Llama 4 Maverick | 60.1 |
| 13 | Gemini 3.1 Pro (Preview) | 53.6 |
| 14 | Grok 3 | 45.8 |
| 15 | Claude Haiku 4.5 | 43.4 |
Interactive version: theaggregate.ai/benchmark?slug=tukabench-refusal-rate · How It Works · Data refreshed daily, snapshot 2026-09-29.