Recent independent benchmarking by researcher Charles AZAM reveals critical performance distinctions between Anthropic Claude Fable 5 and OpenAI GPT-5.6 Sol when tackling complex combinatorial optimization. The evaluation centered on the KIRO problem, an unpublished NP-hard fiber-network routing challenge requiring the connection of over five hundred terminals to distribution hubs using redundant loops and branches while minimizing total cable length. The underlying search space exceeds one thousand two hundred decimal digits of possible configurations, providing a rigorous stress test for autonomous reasoning architectures. AZAM subjected the flagship models to a thirty-minute optimization budget using standardized execution environments. The primary variable under investigation was the native /goal persistence command, designed to extend model attention on complex tasks. Results demonstrate that Fable 5 possesses substantially higher raw optimization capability, consistently outperforming GPT-5.6 Sol across both standard and goal-driven configurations. Fable 5 maintained exceptional consistency, with results fluctuating within a narrow margin, whereas Sol exhibited significantly higher variance and slower convergence. Contrary to expectations, the /goal feature proved to be a double-edged sword for both models. While the persistence mode secured a higher trial win rate across six matched runs, it simultaneously degraded the average performance metrics for both Fable 5 and Sol. The data indicates that /goal fundamentally alters the agent control loop and search trajectory. When models make sound architectural choices, extended persistence allows them to refine and execute superior solutions. However, when early decisions lead to inefficient algorithms or exhaustive search patterns, the mode amplifies these errors by granting additional compute time to flawed strategies. Consequently, the median outcome improved slightly, but the performance tail deteriorated, pulling the mean downward. The investigation further exposes divergent technical implementations of the persistence feature across platforms. Anthropic Claude Code utilizes an independent evaluator model to monitor objective completion based solely on conversation transcripts, limiting its ability to assess file states or iterative solver efficiency. In contrast, OpenAI Codex treats the goal as a persistent thread state stored in local databases, granting the working model direct access to file systems and lifecycle tools. This architectural difference allows Codex to audit its own progress dynamically but introduces the risk of self-reinforcing computational loops. The benchmark underscores a fundamental trade-off in autonomous AI optimization: persistence does not guarantee quality, and extending reasoning time without robust validation mechanisms can degrade average outcomes. AZAM notes that the quality of the underlying decision loop matters more than the duration of the search process. All experimental protocols, source code, execution logs, and analysis scripts have been published in the CLIArena repository to facilitate independent verification. The findings suggest that future AI agents tackling NP-hard problems require more sophisticated early-termination validation and adaptive control architectures rather than simple persistence toggles.