Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?
charlesazam.com
July 18, 2026
7 min read
🔥🔥🔥🔥🔥
60/100
Summary
Fable 5 outperformed GPT-5.6 Sol on an NP-hard optimization problem. The use of the /goal mode did not significantly enhance performance for either model.
Key Takeaways
- Fable 5 outperformed GPT-5.6 Sol in solving an NP-hard optimization problem, providing the best overall solution with high consistency.
- The /goal mode did not significantly enhance performance for either model; it occasionally resulted in worse outcomes despite winning more trials.
- The optimization problem involved designing a fiber network with complex constraints, leading to an enormous search space of approximately 10^1223 possible assignments.
- Both Fable 5 and GPT-5.6 Sol experienced mixed results with the /goal mode, showing minor benefits but also significant regressions in some cases.
Coverage & discussions
Hacker News · charlesazam.com
Community discussions
Hacker News discussion ↗
Historical count: 5 · 5 saved in sample · Historical sample
Analysis ready
Community Sentiment
Positives
- Using /goal has transformed my workflow; it’s like having a mission statement that keeps the model focused and on track.
- Claude's /goal feature feels like giving the model a clear directive, making it operate more effectively during tasks.
- The ability to run parallel investigations in ultra mode is a game-changer for avoiding local optima in complex problems.
- The /goal command is my go-to for ensuring clarity and precision in technical documents — it eliminates vague language.
Concerns
- Claude tends to forget crucial information during long sessions, making /goal feel less effective for extended tasks.
- Compaction in Codex feels like a black box; quality drops off unpredictably, leading to frustrating experiences.
- Using /goal with time constraints seems risky; it might prioritize speed over achieving specific outcomes.
- Ultra mode might be misleading; it’s not a silver bullet and can actually lead to worse results on certain tasks.
