RAI-Bench - Refusal Rate (General): leaderboard
GovTech Singapore's RAI-Bench runs adversarial prompts from RabakBench in four Singapore deployment scenarios, fifty per scenario. This board is the general scenario, where PAIR jailbreak-style prompts are aimed at a general-purpose system prompt. An LLM judge labels each reply Reject, Reframe or Engage, and the number is the share labelled Reject or Reframe, so higher is better.
Metric: Rate (%). Source: huggingface.co. 78 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 84 | #245 |
| 2 | DeepSeek V3.2 | 84 | #205 |
| 3 | GPT-5 Nano | 84 | #425 |
| 4 | GPT-OSS-20B | 84 | #518 |
| 5 | DeepSeek R1 0528 | 84 | #222 |
| 6 | Claude Sonnet 4.5 | 82 | #142 |
| 7 | DeepSeek R1 | 82 | #251 |
| 8 | Llama 4 Maverick | 82 | #464 |
| 9 | Gemini 2.5 Flash Lite | 82 | #422 |
| 10 | Qwen 3.6 Plus | 82 | #120 |
| 11 | Gemma 4 31B | 82 | #200 |
| 12 | Kimi K2 (Thinking) | 82 | #243 (Kimi K2) |
| 13 | Claude 3.5 Haiku | 82 | #568 |
| 14 | Claude Opus 4 | 82 | #163 |
| 15 | GLM-4.5 | 82 | #275 |
Interactive version: theaggregate.ai/benchmark?slug=rai-bench-refusal-rate-general · How It Works · Data refreshed daily, snapshot 2026-10-20.