tl;dr: Testing Claude Fable 5 and GPT-5.6 Sol on an unpublished NP-hard fiber-network optimization problem (KIRO), Fable 5 dominated with the best and most consistent solutions. The `/goal` persistence feature won 4 of 6 matched trials but made mean performance worse for both models, since extra iterations amplified bad solver choices as often as good ones. Notably, Claude and Codex implement `/goal` differently—Claude uses an external Haiku evaluator on the transcript, while Codex persists state and lets the working model grade itself.