I gave Fable 5 and GPT‑5.6 Sol the same unpublished NP‑hard fiber‑network design problem, with and without their native /goal mode. Fable 5 emerged as an absolute beast, delivering the best overall solution and exhibiting unprecedented consistency—an impressive display of raw intelligence.
The /goal feature is not a generic “try harder” switch; it changes the control loop and alters the search path. In some trials it finds a better basin, in others it stalls on a poor idea. Despite winning four of six runs, the mean performance of /goal is actually worse than plain mode, showing that a persistence feature can skew results without improving averages.
The benchmark, called KIRO, asks solvers to connect 532 terminals across Grenoble, Nice, and Paris while respecting loop and branch constraints. Even a simple lower bound yields ~10^1223 possible assignments, underscoring the problem’s enormity. Fable 5’s plain run stayed within a tight 319‑point range, whereas GPT‑5.6 Sol spanned 1,958 points.
All code, prompts, and results are in CLIArena, confirming that Fable 5’s performance is not an artifact of a particular setup.
Source: Hacker News · Summarized by HeadlinesBriefing