Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?
Points and comments are a snapshot, not live.
Fable 5 beats GPT-5.6 Sol on NP-hard problem; /goal mode helps inconsistently.
The author tested Claude Fable 5 and GPT-5.6 Sol on KIRO, an unpublished NP-hard fiber-network optimization problem. Fable 5 produced the best solution and was far more consistent, with plain mean 32,386 vs Sol's 34,261. /goal mode improved median performance slightly but caused occasional large regressions, making both models' means worse. The effect stems from /goal sustaining either good or bad solver strategies. Claude Code and Codex implement /goal differently: Claude uses a separate evaluator (Haiku) that sees only the transcript; Codex persists the goal and lets the working model self-declare completion.
What commenters are saying
Commenters debated /goal's practical utility. Several praised it for long-running tasks, using time-bounded goals to force thorough design docs. Others cautioned that /goal can produce 'pinky promises' of completion without true verification, especially when the model grades its own work. A commenter noted that hypothesis generation is critical: good agents must sample the search space broadly before committing to a solution. One commenter warned that overfitting to stated goals can cause the model to sacrifice unstated priorities, likening it to a paperclip problem.