SQuID: leaderboard
Quantitative spatial reasoning on satellite imagery from SQuID, the Satellite Quantitative Intelligence Dataset (Massih and Cosatto, 2026): 2,000 questions on land-cover percentages, object counts, proximity, connectivity and multi-condition area calculations at 0.3-1.0 m resolution. Accuracy in percent counts an answer inside the range set from human annotation variance, or exact for categorical questions, with the vision-language model answering directly, zero-shot (higher is better).
Metric: Accuracy (%; answer inside the human-variance range or exact for categorical questions, 2,000 questions over three tiers; direct VLM, zero-shot pass@1). Source: arxiv.org. Saturation forecast: Around March 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Minimal) | 28.1 | #163 (GPT-5) |
| 2 | Qwen 3 VL 30B A3B (Thinking) | 26.45 | #710 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=squid · How It Works · Data refreshed daily, snapshot 2026-11-12.