On July 17, 2026, Charles Azam tested Claude Fable 5 and GPT-5.6 Sol on an unpublished NP-hard fiber-network design problem originally created for a hackathon, comparing their performance with and without the native '/goal' mode. Fable 5 delivered the best overall solution and showed remarkable consistency, surpassing GPT-5.6 Sol and human-coded baselines, according to charlesazam.com.
Azam, who previously wrote a C++ solver for this problem, ran both models through the same benchmark over a week. He found that Fable 5 demonstrated superior raw intelligence and solution quality. The '/goal' mode, designed to alter the models' search strategies, did not consistently improve results; it sometimes helped find better solutions but occasionally led to poorer outcomes. All data, code, and prompts are available on the CLIArena GitHub repository linked in the blog post.
This test highlights Fable 5's strength in tackling complex optimization tasks, an area critical for operations research and network design. The results suggest that while control modes like '/goal' can influence model behavior, they are not a universal enhancement. Fable 5's performance sets a new benchmark for AI solving NP-hard problems, exceeding GPT-5.6 Sol's capabilities in this domain.
The full analysis, including score distributions and trajectory notes, is documented on charlesazam.com as of July 17, 2026. The benchmark and related materials remain open for review on the CLIArena GitHub repository, providing a resource for further AI research in optimization challenges.