Atlas

Benchmarks

← All benchmarks

VPCT

Multimodal · 2025-04-19

VPCT contains 100 elementary visual-physics problems in which a ball travels down ramps toward buckets. Models must predict which bucket will catch the ball. Published leaderboard values are mean percentage accuracy across repeated runs of these same 100 problems; per-model run counts are not published.

Top models (higher is better)

ModelScore
Gemini 3 Pro Preview91.0
GPT-5.284.0
Gemini 3 Flash Preview72.6
GPT-566.0
GPT-5.158.7
O4 Mini57.5
o352.0
Gemini 2.5 Pro46.4
Gemini 2.5 Flash Preview (09-2025)46.2
GPT-4.545.0
Gemini Robotics-ER 1.540.8
Gemini 2.5 Pro Preview 05-0640.5
GPT-5 Mini40.2
Opus 4.540.0
GPT-4o (source unspecified)40.0
Sonnet 4.539.8
Claude 3.7 Sonnet39.0
Opus 438.0
Gemini 2.5 Flash Preview 04-1738.0
GPT-5 Nano37.2
o1 (2024-12-17)37.0
Opus 4.135.0
Sonnet 434.0
GPT-4o Mini34.0
GPT-4.133.0
Gemini 2.5 Flash-Lite Preview (09-2025)30.0
Loading Atlas data…