RAI-Bench - Refusal Rate (Career): leaderboard
GovTech Singapore's RAI-Bench runs adversarial prompts from RabakBench in four Singapore deployment scenarios, fifty per scenario. This board is the career advisory chatbot scenario, where the prompts are employment fraud scripts such as fake job offers aimed at Singaporean workers. 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. Saturation forecast: Not forecast. 77 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 100 | #245 |
| 2 | Claude Opus 4.6 | 100 | #63 |
| 3 | Claude Haiku 4.5 | 100 | #280 |
| 4 | Claude Sonnet 4 | 100 | #199 |
| 5 | Gemini 3 Flash (Preview) | 100 | #82 |
| 6 | Gemini 3.5 Flash | 100 | #60 |
| 7 | Gemini 3.1 Flash Lite | 100 | #191 |
| 8 | Gemma 4 31B | 100 | #200 |
| 9 | Gemini 3.6 Flash | 100 | #51 |
| 10 | Claude 3.5 Haiku | 100 | #568 |
| 11 | Gemini 2.5 Pro | 98 | #152 |
| 12 | Gemini 3.1 Pro (Preview) | 98 | #58 |
| 13 | Claude Sonnet 4.6 | 98 | #90 |
| 14 | Claude Opus 4.5 | 98 | #83 |
| 15 | Claude Opus 4 | 98 | #163 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rai-bench-refusal-rate-career · How It Works · Data refreshed daily, snapshot 2026-10-20.