Nejumi 4 - HLE (Japanese) - Calibration Error: leaderboard
Calibration error (%) of the confidence each model states with its answers on hle-ja, the Japanese Humanity's Last Exam subset (194 test questions) in Nejumi Leaderboard 4 (Weights & Biases Japan). Scored as in HLE's grader; 0 is perfectly calibrated and lower is better.
Metric: Calibration error (%). Source: nejumi.ai. Saturation forecast: Around March 2027. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.5 (High) | 34 | #54 (Grok 4.5) |
| 2 | Grok 4.20 0309 (Reasoning) | 39 | #120 (Grok 4.20 0309) |
| 3 | Claude Opus 4.8 (xHigh) | 39 | #39 (Claude Opus 4.8) |
| 4 | Claude Opus 5 (Max) | 41 | #16 (Claude Opus 5) |
| 5 | Muse Spark Open (xHigh) | 41 | #42 (Muse Spark Open) |
| 6 | Claude Opus 4.7 (xHigh) | 42 | #51 (Claude Opus 4.7) |
| 7 | GPT-5.6 Sol (Max) | 44 | #22 (GPT-5.6 Sol) |
| 8 | GPT-5.4 (2026-03-05) (xHigh) | 44 | #63 (GPT-5.4 (2026-03-05)) |
| 9 | GPT-5.4 (2026-03-05) (High) | 44 | #63 (GPT-5.4 (2026-03-05)) |
| 10 | GPT-5.6 Terra (Max) | 48 | #40 (GPT-5.6 Terra) |
| 11 | Kimi K3 (Thinking) | 48 | #35 (Kimi K3) |
| 12 | Hy4 preview (High) | 49 | #45 (Hy4 preview) |
| 13 | DeepSeek V4.1 Flash | 51 | #37 |
| 14 | GLM-5.3 (Max) | 52 | #44 (GLM-5.3) |
| 15 | DeepSeek-V4-Flash-Vision-Exp (Max) | 53 | #88 (DeepSeek-V4-Flash-Vision-Exp) |
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-calibration-error · How It Works · Data refreshed daily, snapshot 2026-11-05.