Nejumi 4 - HLE (Japanese) - Accuracy: leaderboard
hle-ja, a Japanese translation of Humanity's Last Exam, on the 194-question test subset used by Nejumi Leaderboard 4 (Weights & Biases Japan). Answers are graded by o3-mini; accuracy (%), higher is better.
Metric: Accuracy (%). Source: nejumi.ai. Saturation forecast: Around March 2027. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 52.06 | #22 (GPT-5.6 Sol) |
| 2 | Claude Opus 5 (Max) | 50.52 | #16 (Claude Opus 5) |
| 3 | Claude Opus 4.8 (xHigh) | 48.97 | #39 (Claude Opus 4.8) |
| 4 | GPT-5.6 Terra (Max) | 47.42 | #40 (GPT-5.6 Terra) |
| 5 | GPT-5.4 (2026-03-05) (xHigh) | 45.88 | #63 (GPT-5.4 (2026-03-05)) |
| 6 | Claude Opus 4.7 (xHigh) | 44.33 | #51 (Claude Opus 4.7) |
| 7 | Kimi K3 (Thinking) | 44.33 | #35 (Kimi K3) |
| 8 | Hy4 preview (High) | 44.33 | #45 (Hy4 preview) |
| 9 | Gemini 3.1 Pro (Preview) | 43.81 | #64 |
| 10 | DeepSeek V4.1 Flash | 43.81 | #37 |
| 11 | GPT-5.6 Luna (Max) | 42.78 | #79 (GPT-5.6 Luna) |
| 12 | Grok 4.5 (High) | 42.78 | #54 (Grok 4.5) |
| 13 | GPT-5.4 (2026-03-05) (High) | 41.75 | #63 (GPT-5.4 (2026-03-05)) |
| 14 | Muse Spark Open (xHigh) | 41.24 | #42 (Muse Spark Open) |
| 15 | Kimi K3 (Max) | 38.14 | #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-hle-japanese-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.