Fable 5 vs GPT-5.6 Sol: NP-Hard Problem Benchmark Results
Original: Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?
Why This Matters
Real-world NP-hard benchmarks offer a critical stress test for comparing frontier AI coding agents beyond standard benchmarks.
Engineer Charles Azam tested Claude Fable 5 and GPT-5.6 Sol on an unpublished NP-hard fiber-network optimization problem with 30-minute budgets, with and without /goal mode. Fable 5 produced the best overall solution and showed superior consistency across three matched runs.
Charles Azam benchmarked Claude Fable 5, Opus 4.8, Sonnet 5, GPT-5.6 Sol, Terra, and Luna on KIRO — a fiber-network design problem from a 2018 engineering hackathon involving directed distance matrices for Grenoble, Nice, and Paris. The solver must minimize total cable length while connecting distribution hubs and terminals via redundant loops and short branches under strict structural constraints. The search space is estimated at roughly 10^1223 possible configurations for Paris alone.
Each model was given a 30-minute optimization budget with maximum reasoning settings via Harbor 0.1.43 running Docker. Three matched runs were conducted for Fable 5 and GPT-5.6 Sol, each tested both in plain mode and with the native /goal command.
Fable 5 achieved the best individual result (31,934 in Run 1 with /goal) and demonstrated notably higher consistency. GPT-5.6 Sol showed wider variance, with /goal sometimes degrading performance significantly (+5,790 in Run 1). Across six trials, /goal won four, but the author cautions that win rate alone is misleading — /goal modifies the control loop and search path differently per model, acting as a separate evaluator in Claude Code and a stateful lifecycle tool in Codex, rather than a simple 'try harder' switch.