BenchCompass - Context-Grounded Reasoning (Open Book): leaderboard
BenchCompass Pro: the same 96 payment scenarios with the evidence pack supplied; the closed-to-open gap separates poor evidence use from missing knowledge. Higher is better. Source: arxiv.org/abs/2609.18270.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 89.6 | #64 |
| 2 | Qwen 3.5 397B A17B | 84.4 | #154 |
| 3 | Qwen 3.5 122B A10B | 84.4 | #184 |
| 4 | GPT-5.5 | 82.3 | #33 |
| 5 | Claude Opus 4.7 | 82.3 | #51 |
| 6 | Gemma 4 31B (IT) | 82.3 | #178 |
| 7 | Gemma 4 26B A4B (IT) | 82.3 | #293 |
| 8 | DeepSeek V4 Pro | 81.3 | #112 |
| 9 | MiniMax-M2.5 | 75 | #316 |
| 10 | GLM-5.1 | 74 | #109 |
| 11 | Qwen 3 32B | 69.8 | #449 |
| 12 | Llama 4 Maverick | 68.8 | #486 |
| 13 | Mistral Large 3 | 64.6 | #410 |
| 14 | Qwen 3 8B | 58.3 | #678 |
| 15 | Ministral 3 14B | 57.3 | #664 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=benchcompass-context-grounded-reasoning-open-book · How It Works · Data refreshed daily, snapshot 2026-11-05.