Fable 5 beats GPT-5.6 Sol on NP-hard KIRO: 5.5% Gap, six times more robust
Fable 5 outperformed GPT-5.6 Sol on the NP-Hard KIRO problem by 5.5%. It scored 32,386 on average vs. 34,261 for GPT-5.6 Sol, with six times smaller variance. /goal mode helped win more runs but lowered the overall score due to agentic behavior.
Charles Azam, developer, compared Fable 5 and GPT-5.6 Sol on the KIRO problem (a French hackathon). It’s about optimizing the length of fiber optics in a network (NP-Hard). There are no known exact solutions except brute force. Both models were run for 30 min and had 3 tries with maximal reasoning.
Results: Fable 5 Is More Accurate and Six Times More Consistent
Fable 5 scored on average 32,386 units (the lower, the better, it’s a minimization task) vs. 34,261 for GPT-5.6 Sol, i.e., 5.5% difference. Fable 5 had a variance of 319 units while Sol had 1,958 — so Sol is 6 times less reliable. Fable 5 got the best score of 31,934 units.
The /goal Mode Paradox
/goal mode allows to persist until the deadline. This means that the model continues searching for better solutions until the end. It helped to win 4 out of 6 runs, but ironically decreased the overall score because of agentic behavior. “Persistence” and “solution quality” may be conflicting goals for agentic systems.
Source: charlesazam.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →
Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles