Nejumi 4 - Toxicity - Social Norms: leaderboard
Japanese safety prompts on social norms (category 社会規範) 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 | GPT-5 Nano (2025-08-07) (High) | 99.6 | |
| 2 | GPT-5.2 (2025-12-11) (xHigh) | 99.6 | |
| 3 | Qwen 3.6 27B (Thinking) | 99.6 | #182 (Qwen 3.6 27B) |
| 4 | Claude Opus 4.8 (xHigh) | 99.6 | #39 (Claude Opus 4.8) |
| 5 | GPT-5.4 (2026-03-05) (High) | 99.6 | #63 (GPT-5.4 (2026-03-05)) |
| 6 | DeepSeek V3.2 (Thinking) | 99.21 | #210 (DeepSeek V3.2) |
| 7 | GPT-5.4 Mini (2026-03-17) (High) | 99.21 | #180 (GPT-5.4 Mini (2026-03-17)) |
| 8 | Claude Sonnet 4 (Non-reasoning) | 99.21 | #207 (Claude Sonnet 4) |
| 9 | GPT-5.6 Luna (Max) | 98.81 | #79 (GPT-5.6 Luna) |
| 10 | ELYZA-Shortcut-1.0-Qwen-32B | 98.81 | #560 |
| 11 | Qwen 3.6 35B A3B (Thinking) | 98.81 | #238 (Qwen 3.6 35B A3B) |
| 12 | K-EXAONE (Thinking) | 98.81 | #387 (K-EXAONE) |
| 13 | Qwen 3.6 Max (Preview) (Thinking) | 98.81 | |
| 14 | GPT-5.6 Terra (Max) | 98.41 | #40 (GPT-5.6 Terra) |
| 15 | Claude Sonnet 4.6 (Thinking) | 98.41 | #94 (Claude Sonnet 4.6) |
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-social-norms · How It Works · Data refreshed daily, snapshot 2026-11-05.