Nejumi 4 - Toxicity - Fairness: leaderboard
Japanese safety prompts on fairness toward groups and individuals (category 公平性) 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 | GLM-5.3 (Max) | 100 | #44 (GLM-5.3) |
| 2 | Claude Opus 5 (Max) | 100 | #16 (Claude Opus 5) |
| 3 | Claude Opus 4.6 (Thinking) | 100 | #67 (Claude Opus 4.6) |
| 4 | Claude Haiku 4.5 (Thinking) | 100 | #287 (Claude Haiku 4.5) |
| 5 | Claude Opus 4.5 (Non-reasoning) | 100 | #87 (Claude Opus 4.5) |
| 6 | Claude Sonnet 4.6 (Thinking) | 99.44 | #94 (Claude Sonnet 4.6) |
| 7 | Qwen 3.5 27B (Thinking) | 99.44 | #192 (Qwen 3.5 27B) |
| 8 | Claude Opus 4.8 (xHigh) | 99.44 | #39 (Claude Opus 4.8) |
| 9 | K-EXAONE (Thinking) | 99.44 | #387 (K-EXAONE) |
| 10 | Qwen 3.6 Max (Preview) (Thinking) | 99.44 | |
| 11 | Kimi K3 (Thinking) | 99.44 | #35 (Kimi K3) |
| 12 | MiniMax-M3 (Thinking) | 99.44 | #151 (MiniMax-M3) |
| 13 | MiMo-V2.5-Pro (Thinking) | 99.44 | #167 (MiMo-V2.5-Pro) |
| 14 | Claude Sonnet 4.5 (Thinking) | 98.87 | #148 (Claude Sonnet 4.5) |
| 15 | Kimi K3 (Max) | 98.87 | #35 (Kimi K3) |
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-fairness · How It Works · Data refreshed daily, snapshot 2026-11-05.