A recent evaluation pitted OpenAI's GPT-5.6 family against Anthropic's Claude Fable 5 on five physical modeling and simulation problems. The goal was to determine which frontier model excels at accurately representing real-world physics within AI agents.

Claude Fable 5 achieved the highest weighted score of 0.889, despite being the most expensive at $9.60 per trial. GPT-5.6-sol followed with a score of 0.814 and a cost of $1.74 per trial. GPT-5.6-terra scored 0.786 at $1.25 per trial, while GPT-5.6-luna had the lowest score of 0.727 and the longest run time.

Each model exhibited distinct work styles. Fable meticulously verified its solutions, Sol over-specified tasks, Luna iterated excessively, and Terra economized by reducing simulation horizons. None of the models successfully solved the most complex problem, modeling NASA's HL-20 flight vehicle.