FMG-Bench - Guided Harness: leaderboard
FMG-Bench v1 (arXiv 2608.12324v1 Table 8; authors' results/model_scores.csv): 14 models on 628 rendered theological and pastoral scenarios across four triage levels, scored 0-100 by a three-model judge panel (GPT-5.4 mini, Gemini 3.1 Flash-Lite Preview, Claude Sonnet 4.6); higher is better. Guided-default condition: a structured harness adds triage rules, grounding discipline and escalation boundaries.
Metric: Judge-panel score (0-100; guided default harness). Source: github.com. Saturation forecast: Estimated already saturated. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 94.58 | #51 |
| 2 | Kimi K2.6 | 94.14 | #113 |
| 3 | GPT-5.4 | 93.9 | #83 |
| 4 | Grok 4.20 | 93.85 | #179 |
| 5 | Qwen 3.6 Plus | 93.83 | #124 |
| 6 | DeepSeek V4 Pro | 92.81 | #112 |
| 7 | GLM-5.1 | 92.65 | #109 |
| 8 | Gemini 3.1 Pro (Preview) | 91.68 | #64 |
| 9 | Nemotron 3 Super | 90.99 | #344 |
| 10 | MiMo-V2.5-Pro | 90.93 | #167 |
| 11 | Mistral Large 3 | 89.92 | #410 |
| 12 | MiniMax-M2.7 | 88.95 | #270 |
| 13 | Seed 2.0 Lite | 88.35 | #161 |
| 14 | Llama 4 Maverick | 79.3 | #486 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=fmg-bench-guided-harness · How It Works · Data refreshed daily, snapshot 2026-11-05.