Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help? (charlesazam.com)

A recent experiment tested Claude Fable 5 and GPT-5.6 Sol on an NP-hard optimization problem called KIRO, originally developed for a hackathon. The goal was to optimize a fiber-network design problem across various cities while minimizing cable length. Fable 5 emerged as an incredibly robust performer, consistently producing superior solutions compared to GPT-5.6 Sol, even without its native /goal setting, which did not significantly enhance its performance. The results indicated that while the /goal feature showed some promise in specific instances, it often led to regressions in average results for both models. Significantly, this study underscores the complexities of optimization tasks in AI, revealing that features designed to enhance performance can sometimes yield counterintuitive results. The findings suggest that a model's ability to maintain effective control loops and decision-making pathways is crucial, especially in challenging optimization scenarios where a misstep can exacerbate performance issues. These insights are vital for the AI/ML community, highlighting the need for careful evaluation of system features and optimization approaches, especially in high-dimensional and NP-hard problem settings.