ReasonScape M12X: leaderboard
12 cognitive domains with 3-degree difficulty scaling. Parametric generators create difficulty manifolds to pinpoint exactly where models break. 100+ models evaluated on 9B+ tokens. Open source.
Metric: ReasonScore (hard). Source: reasonscape.com. Saturation forecast: Around December 2026. 81 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-4.5 | 529.8 | #275 |
| 2 | Qwen 3 Next 80B A3B (Thinking) | 481.4 | #320 (Qwen 3 Next 80B A3B) |
| 3 | Qwen 3 235B A22B FP8 | 473.79 | #301 |
| 4 | Hunyuan A13B-Instruct | 467.03 | #666 |
| 5 | Qwen 3 VL 32B (Thinking) | 410.44 | #294 (Qwen 3 VL 32B) |
| 6 | Qwen 3 VL 8B (Thinking) | 299.15 | |
| 7 | Gemma 3 27B (IT) | 260.24 | #523 |
| 8 | Gemma 3 12B (IT) | 196.49 | #664 |
| 9 | Qwen 3 VL 4B (Thinking) | 191.77 | #480 (Qwen 3 VL 4B) |
| 10 | OLMo 3 7B (Thinking) | 151.9 | #1112 (OLMo 3 7B) |
| 11 | ERNIE-4.5-21B-A3B (Thinking) | 106.43 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=reasonscape-m12x · How It Works · Data refreshed daily, snapshot 2026-10-20.