BenchCompass - Context-Grounded Reasoning (Closed Book): leaderboard
BenchCompass Pro: 96 open-QA payment scenarios (transaction state, participant role, region, rail) answered without the evidence pack, so success means the rules were already known. Higher is better. Source: arxiv.org/abs/2609.18270.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around September 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 36.5 | #33 |
| 2 | Claude Opus 4.7 | 25 | #51 |
| 3 | DeepSeek V4 Pro | 16.7 | #112 |
| 4 | GLM-5.1 | 15.6 | #109 |
| 5 | Gemini 3.1 Pro (Preview) | 14.6 | #64 |
| 6 | Qwen 3.5 397B A17B | 14.6 | #154 |
| 7 | Gemma 4 31B (IT) | 9.4 | #178 |
| 8 | Qwen 3.5 122B A10B | 9.4 | #184 |
| 9 | MiniMax-M2.5 | 9.4 | #316 |
| 10 | Llama 4 Maverick | 8.3 | #486 |
| 11 | Mistral Large 3 | 8.3 | #410 |
| 12 | Gemma 4 26B A4B (IT) | 7.3 | #293 |
| 13 | Qwen 3 32B | 5.2 | #449 |
| 14 | Ministral 3 14B | 2.1 | #664 |
| 15 | Ministral 3 8B | 2.1 | #699 |
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-closed-book · How It Works · Data refreshed daily, snapshot 2026-11-05.