Fable 5 vs GPT-5.6 Sol on an NP-Hard Problem: Does /goal Help?
Claude Fable 5 and GPT-5.6 Sol tackle an NP-hard fiber-network problem, testing whether the /goal persistence feature actually improves results.
An unpublished NP-hard optimization task called KIRO, a fiber-network design problem with a search space far larger than typical coding benchmarks, was used to compare Claude Fable 5 against GPT-5.6 Sol, both in plain mode and using each CLI's native /goal persistence feature. Fable 5 produced the best overall solution and showed markedly tighter consistency across runs than Sol, a clear raw-capability gap.
The more counterintuitive finding concerns /goal itself: it won four of six matched trials for both models, yet worsened average performance for each. The feature occasionally delivered a modest gain but sometimes let a bad solver strategy run far longer than it should have, producing large regressions that dragged down the mean even as the median improved slightly.
Digging into implementations shows why: Claude Code's /goal relies on a separate evaluator model (Haiku by default) that judges only the transcript with no tool or file access, while Codex persists goal state and lets the working model itself declare and grade completion. The practical takeaway for engineers building agentic workflows is that a persistence mechanism can win most individual trials while making observed average performance worse, so what the loop keeps doing matters more than whether it keeps going.