GPT-5.6 vs. Claude Fable 5 for Physical AI, which performs best?

Claude Fable 5 outperformed GPT-5.6 variants in physical AI modeling and simulation tasks, achieving a weighted score of 0.889 compared to GPT-5.6-terra's 0.786, according to JuliaHub's internal evaluation from July 2026. The testing involved five problems of varying difficulty where agents derived physics models, compiled them, and simulated results against ground truth benchmarks. Claude Fable 5 also demonstrated cost efficiency at $9.60 per trial despite slightly longer execution times.