Fable 5 vs. GPT-5.6: The Risks of AI Persistence in Optimization

A benchmark of Claude Fable 5 and GPT-5.6 Sol on an NP-hard optimization problem shows that Fable 5 is the superior model in terms of intelligence and consistency. The study reveals that the native '/goal' feature is a double-edged sword that can actually degrade average performance by encouraging models to persist in poor strategies. Ultimately, the research suggests that a model's inherent reasoning capability is more critical than persistence-based control loops.

Key Points

Claude Fable 5 outperformed GPT-5.6 Sol in both raw optimization scores and consistency on the NP-hard KIRO benchmark.
The '/goal' feature improved median performance but worsened mean performance because it can amplify bad strategies just as easily as good ones.
Claude Code and OpenAI's Codex use different architectures for goal persistence: Claude uses a separate evaluator model, while Codex uses persisted thread state and specific tools.
On complex optimization tasks, the quality of the model's initial reasoning and strategy is more important than the persistence of the execution loop.

Sentiment

The overall sentiment is cautiously positive toward the article's core point. Hacker News largely accepts that goal persistence and stronger agent harnesses are useful but risky, and many comments add practical anecdotes that support the article's concerns. At the same time, the community is not fully convinced by the benchmark as decisive evidence; the tone is constructive, skeptical, and focused on methodology, baseline comparisons, transparency, and cost.

In Agreement

The article's warning about persistence resonated with users who have seen high-effort agent modes overthink tasks, spend extra time, and pursue flawed approaches instead of adapting.
Several commenters agreed that harness behavior matters as much as the underlying model, especially when a goal loop or multi-agent scaffold changes the search strategy.
Users supported the idea that Fable can be more insightful for tasks requiring intent inference, judgment, and broad reasoning, while Sol's relentlessness can produce both benchmark strength and practical risk.
The thread reinforced the need to evaluate goal-oriented modes per workload rather than treating them as universal upgrades.
Comments about long sessions, compaction, and protected goal messages aligned with the article's broader concern that persistent context and objective tracking can preserve useful intent while also creating brittle behavior.

Opposed

Some commenters argued that the benchmark evidence is too noisy or too limited to justify strong model-level conclusions.
A few pushed back on the fairness of the comparison, including whether the OpenAI model was evaluated under equivalent conditions or with the most appropriate mode.
Several argued that hard combinatorial optimization should be compared against conventional solvers, integer programming formulations, or specialized search systems rather than only against LLM agents.
Some commenters suggested the final result is less informative than the trajectory of improvement over time, since persistence may look different if measured by best-so-far progress.
A minority were broadly dismissive of both flagship models for deep complex work, arguing that they degrade as tasks become more specialized.