Nejumi 4 - ALT - Robustness: leaderboard
Each JMMLU multiple-choice question is put three ways: the standard form, a form with symbols as answer labels, and one asking for all the incorrect options. An item scores one point when all three replies point to the same option, half a point when two do and nothing when all differ. The board is that consistency rate in percent from the few-shot run, higher is better. Source: github.com/wandb/llm-leaderboard.
Metric: Score (%). Source: nejumi.ai. Saturation forecast: Estimated already saturated. 139 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 100 | #58 |
| 2 | O3 (2025-04-16) | 100 | #121 |
| 3 | GPT-5.6 Sol | 99.5 | #19 |
| 4 | Claude Opus 5 | 99 | #12 |
| 5 | GPT-5.6 Terra | 99 | #41 |
| 6 | GPT-5.5 (2026-04-23) | 99 | #22 |
| 7 | Gemini 2.5 Pro | 98.5 | #152 |
| 8 | Claude Opus 4.7 | 98 | #49 |
| 9 | Gemini 3.6 Flash | 98 | #51 |
| 10 | Qwen 3.5 27B | 97.5 | #187 |
| 11 | Gemini 3.5 Flash | 97.5 | #60 |
| 12 | Claude Fable 5 | 97.5 | #16 |
| 13 | DeepSeek V4.1 Flash | 97.5 | #36 |
| 14 | Qwen 3.6 Max Preview | 97.5 | #85 |
| 15 | Qwen3.8 2.4T A95B | 97.5 | #55 |
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-alt-robustness · How It Works · Data refreshed daily, snapshot 2026-10-20.