- Claude Fable 5 outperformed GPT-5.6 Sol on an NP-hard optimization problem, demonstrating superior consistency and producing the best overall solution.
- The /goal mode had mixed results; it improved outcomes in some runs but caused significant regressions in others, leading to worse average performance despite winning most individual trials.
- The study highlights that /goal is not a generic 'try harder' switch but alters the control loop and search path, which can amplify both good and bad decisions.