RedVox: leaderboard
RedVox (Savoldi et al., 2026) tests speech and omni models with harmful and stereotyped requests recorded by native speakers, adapted from M-ALERT safety prompts and SHADES stereotypes; each request is either spoken or written with a silent or noisy audio clip attached. A GPT-5.5 judge labels each response safe, safe by misunderstanding, controversial or unsafe; the score is the unsafe share in percent over the five languages of the released set, lower is better.
Metric: Unsafe response rate (%; share of responses a GPT-5.5 judge labels unsafe (endorsing, supporting or complying with the harmful or stereotyped request, whether or not the request was understood); all five languages and request types of the 3,414 released entries). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Flash Lite | 2.6 | #363 |
| 2 | Gemini 3.1 Pro (Preview) | 3.1 | #91 |
| 3 | Qwen3 Omni 30B A3B Instruct | 3.4 | #902 |
| 4 | Phi-4 Multimodal Instruct | 16.1 | #3914 |
| 5 | Voxtral-Small-24B-2507 | 21.9 | #1291 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redvox · How It Works · Data refreshed daily, snapshot 2026-11-12.