VPCT

What does it measure?
Tests intuitive physics: predicting which bucket a ball will land in after rolling down ramps.
Who evaluates it, and how?
100 problems, each a picture of a ball, several ramps, and three buckets. Random guessing gets one in three, and the human volunteers all scored 100%.
Reading the score
Scores run from 0 to 100%. Higher is better.
Top score72.6%
Models tested10
Scoring basisHigher is betterImage-input models only
Model release date

Can it predict how objects will move just from a picture?

VPCT

  1. 🥇GoogleGemini 3 Flash72.6%
  2. 🥈OpenAIGPT-5Reasoning effort: High66.0%
  3. 🥉OpenAIGPT-5 MiniReasoning effort: Medium40.2%
  4. #4AnthropicClaude Opus 4.5Reasoning effort: Budget 32K40.0%
  5. #5AnthropicClaude Sonnet 4.5Reasoning effort: Budget 32K39.8%
  6. #6AnthropicClaude Opus 4Reasoning effort: Budget 16K38.0%
  7. #7OpenAIGPT-5 NanoReasoning effort: High37.2%
  8. #8AnthropicClaude Opus 4.135.0%
  9. #9AnthropicClaude Sonnet 4Reasoning effort: Budget 32K34.0%
  10. #10GoogleGemini 2.5 Flash Lite30.0%
Half of models ≤ 38.9%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.