PlanBench-V: leaderboard
Metric: Overall mean score (out of 2) on a 300-item stratified subset of the eight tasks; answers to expert-curated questions on Chinese territorial spatial planning maps are scored 0 to 2 by a GPT-4o-mini judge (temperature 0) against structured reference answers with annotated critical points; objective items use exact match or semantic similarity; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 17 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.6 Plus | 1.7 |
| 2 | Gemini 2.5 Pro | 1.47 |
| 3 | GPT-5.4 | 1.43 |
| 4 | Kimi K2.6 | 1.42 |
| 5 | Claude Opus 4.7 | 1.38 |
| 6 | Qwen 3.6 Flash | 1.35 |
| 7 | GPT-4o | 1.34 |
| 8 | Qwen 2.5 VL 7B Instruct | 1.05 |
| 9 | InternVL3-14B | 0.98 |
| 10 | Qwen 2 VL 7B | 0.91 |
| 11 | InternVL3-8B | 0.91 |
| 12 | GPT-4o Mini | 0.87 |
| 13 | Qwen 2 VL 2B | 0.73 |
Interactive version: theaggregate.ai/benchmark?slug=planbench-v · How It Works · Data refreshed daily, snapshot 2026-09-29.