Atlas

Benchmarks

← All benchmarks

Blueprint-Bench

Multimodal · 2025-09-24

Blueprint-Bench evaluates spatial reasoning by asking models and agents to reconstruct 2D apartment floor plans from approximately 20 photographs per apartment. The score is a connectivity-based similarity measure across 50 apartments.

Top models (higher is better)

ModelScore
Human (Blueprint-Bench)0.5
GPT-50.4
Gemini 2.5 Pro0.4
GPT-5 Mini0.4
Grok 40.4
Gemini 2.5 Flash0.4
Opus 40.4
GPT Image 10.3
Random baseline (Blueprint-Bench)0.3
Sonnet 40.3
Gemini 2.5 Flash Image Preview0.2
GPT-4o (source unspecified)0.1
Loading Atlas data…