Fable 5 vs GPT-5.6 Sol on an NP-Hard Problem: Does /goal Mode Help

We ran a comparative experiment between Claude Fable 5 and GPT-5.6 Sol using the KIRO benchmark, an unpublished NP-hard optimization problem. KIRO is a fiber-optic network design problem, and due to its enormous number of variables, its search space reaches $10^{1223}$, making it extremely complex.

The results showed that Fable 5 delivered overwhelming performance. It not only produced the best solutions but also demonstrated unprecedented consistency on this problem, proving the sheer strength of pure intelligence.

Meanwhile, regarding /goal mode, which changes the model's control loop, we reached the following conclusions.

• /goal mode is not simply a switch that makes the model 'try harder.'
• This mode changes the exploration path and control loop, sometimes leading to better solutions, but sometimes wasting time on flawed ideas.
• As a result, we confirmed that /goal mode is not a universal solution that improves performance in every situation.