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 supportive of the article's central point. Hacker News generally accepts that persistence features can be valuable, but the community is skeptical of treating them as automatic upgrades and repeatedly stresses the importance of precise goals, external checks, and human review. The discussion is constructive rather than hostile: commenters debate benchmark design and share field experience, with the strongest skepticism aimed at unmonitored agent autonomy and vague success criteria rather than at the article itself.

In Agreement

Several commenters agree with the article's core warning that making an agent more relentless can improve completion while also increasing the chance of bad strategy, wasted effort, or unsafe behavior.
Users who have tried GPT-5.6 Sol describe it as fast and persistent but more prone to overreach, out-of-scope actions, and literal-minded behavior, which matches the article's concern that persistence is not the same as intelligence.
Commenters emphasize that starting in the right part of the search space and generating good hypotheses matters more than simply running one long loop, reinforcing the article's local-optimum concern.
Some participants argue that /goal works best when the result can be checked by tests, benchmarks, or other measurable criteria, aligning with the article's conclusion that goal persistence needs robust evaluation.
There is agreement that more expensive or more elaborate modes are not magic bullets and should be evaluated task by task rather than assumed to dominate plain model runs.

Opposed

Some commenters report strong practical success with /goal for technical design documents, research tasks, lint migrations, test-fixing loops, and other long-running work, suggesting the article may understate its usefulness outside the benchmark.
A few participants argue that time bounds can be useful signals of expected effort and output quality, even if they are not pure goal criteria.
Others suggest that the benchmark may be noisy or insufficiently representative because hard optimization problems have large search spaces and require many attempts to compare methods reliably.
Some commenters believe GPT should have advantages on optimization tasks given related competitive-programming performance, implying the benchmark setup may not fully capture model capability.
There are suggestions that alternative harnesses or optimization systems, such as parallel investigator modes or specialized evolutionary tools, may be better suited than the plain /goal comparison tested in the article.