The performance of Fable 5 and GPT-5.6 Sol has been compared on an NP-hard optimization problem, known as the KIRO fiber-network design problem [1]. The results show that Fable 5 outperforms GPT-5.6 Sol, with and without the /goal feature. The /goal feature, which is designed to help the models focus on the objective, has mixed results, sometimes improving and sometimes worsening the outcome [2]. The study found that Fable 5 was able to produce the best solution overall, with a total cable length of 31,934, while GPT-5.6 Sol had a total cable length of 33,581 [3].
The /goal feature was found to be not a generic 'try harder' switch, but rather a feature that changes the control loop and the search path [4]. The study concludes that while the /goal feature can be useful in some cases, it is not a guarantee of better performance, and the quality of the loop matters less than the quality of what the loop keeps doing [5].
Sources
- Charles Azam, 'Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal Help?'
- Anthropic’s goal documentation
- Codex CLI 0.144.4
- Claude Code and Codex implementations of
/goal - CLIArena benchmark task, wrappers, analysis scripts, figure generator, and full evidence memo
