Nejumi 4 - Toxicity - Prohibited Acts: leaderboard
Japanese safety prompts on requests for prohibited or harmful acts (category 禁止行為) from the toxicity suite of Nejumi Leaderboard 4 (Weights & Biases Japan). GPT-4.1 checks each response against category-specific binary criteria; score is the share of criteria met (%), higher is safer.
Metric: Criteria met (%). Source: nejumi.ai. Saturation forecast: Estimated already saturated. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B (Thinking) | 100 | #154 (Qwen 3.5 397B A17B) |
| 2 | Qwen 3.5 27B (Thinking) | 99.48 | #192 (Qwen 3.5 27B) |
| 3 | Qwen 3.6 Max (Preview) (Thinking) | 99.48 | |
| 4 | GPT-5.2 (2025-12-11) (xHigh) | 98.44 | |
| 5 | Qwen 3.6 27B (Thinking) | 98.44 | #182 (Qwen 3.6 27B) |
| 6 | Qwen 3.5 122B A10B (Thinking) | 98.44 | #184 (Qwen 3.5 122B A10B) |
| 7 | Claude Opus 4.8 (xHigh) | 98.44 | #39 (Claude Opus 4.8) |
| 8 | GPT-5.1 (2025-11-13) (Non-reasoning) | 97.92 | #107 (GPT-5.1 (2025-11-13)) |
| 9 | Nova 2.0 Lite (High) | 97.92 | #425 (Nova 2.0 Lite) |
| 10 | DeepSeek V3.2 (Thinking) | 97.4 | #210 (DeepSeek V3.2) |
| 11 | GPT-5.6 Terra (Max) | 97.4 | #40 (GPT-5.6 Terra) |
| 12 | GPT-5 Nano (2025-08-07) (High) | 97.4 | |
| 13 | Claude Opus 4.5 (Non-reasoning) | 97.4 | #87 (Claude Opus 4.5) |
| 14 | Qwen 3.6 35B A3B (Thinking) | 97.4 | #238 (Qwen 3.6 35B A3B) |
| 15 | GPT-5.4 (2026-03-05) (High) | 97.4 | #63 (GPT-5.4 (2026-03-05)) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-toxicity-prohibited-acts · How It Works · Data refreshed daily, snapshot 2026-11-05.