Gen-4.5 Tops MidcenturyAI Physics Bench with 30% Score

The benchmark measures "Time To Failure" by providing models with a single frame from a 1 second simulated physics clip and a text description, then requiring them to generate the following 2 seconds. Failure occurs when a physical violation—such as objects vanishing or passing through each other—persists for at least 0.3 seconds. Results are averaged over 50 identical episodes per scene to ensure a distribution of failures rather than single lucky outputs.