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

A recent study pitted cutting-edge AI models, specifically Claude Fable 5 and GPT-5.6 Sol, against an NP-hard optimization problem to assess their raw intelligence and the impact of their native /goal modes. The problem, dubbed KIRO, is a fiber-network design challenge involving connecting distribution points and terminals using loops and branches while minimizing total cable length. Its search space is astronomically vast, estimated at 10^1223 even for a restricted family of solutions.

• The Problem: KIRO requires designing redundant fiber loops rooted at hubs with short branches, ensuring every tower appears once and accounting for directional cable costs.
• Models and Modes: The primary comparison focused on Claude Fable 5 and GPT-5.6 Sol in both plain and /goal modes, with other models (Opus, Sonnet, Terra, Luna) also briefly evaluated.
• Fable 5's Dominance: Fable 5 was dubbed an "absolute beast," consistently outperforming GPT-5.6 Sol with greater precision and a much tighter score range, showcasing what the author called "pure raw intelligence."
• /goal Mode Nuance: The /goal feature was not a straightforward performance enhancer. While it led to better results in a majority of individual runs, its impact on the average score was negative, indicating that it could sometimes lead to significant regressions by amplifying poor initial decisions.
• Implementation Differences: The article detailed how Claude Code's /goal uses a separate, smaller evaluator model to check progress, whereas Codex (GPT-5.6 Sol's underlying system) integrates create_goal, get_goal, and update_goal tools directly into the working model's state.
• Limitations: The study was confined to a single, unpublished NP-hard task, and some comparisons involved varying conditions. Hardware factors, such as exposed CPUs, may have favored models with parallel processing capabilities.

Ultimately, the study concluded that on hard optimization tasks, the core quality of the model's problem-solving loop is more critical than a persistence feature like /goal, which can paradoxically improve individual trials while worsening overall average performance.