Fable5とGPT-5.6のNP困難問題比較

TL;DR: I gave Claude Fable 5 and GPT-5.6 Sol the same unpublished NP-hard

optimization problem, with and without their native /goal mode. Fable 5 is an absolute beast;

/goal is not a game changer.

Context: This is an operations research problem originally submitted to students at a

hackathon. I spent a week years

ago writing C++ to solve it, so I have a useful human baseline.

Fable 5 was an absolute beast on this benchmark. It produced the best solution overall, and its

consistency is unlike anything I have seen from a model on this problem. This is pure raw

intelligence. Incredible.

The other result is that /goal is not a generic “try harder” switch. It changes the control

loop and the search path. Sometimes that finds a better basin. Sometimes it gives a bad idea

more time to mature.

All code, prompts, result tables, exclusions, and trajectory notes are in

CLIArena. This is a follow-up to my

first article about this benchmark.

The problem

KIRO is a fiber-network design problem I worked on as an engineering student in 2018. Given

directed distance matrices for Grenoble, Nice, and Paris, the solver has to connect distribution

points and terminals using loops and short chains while respecting several structural

constraints. The objective is total cable length. Lower is better.

A valid network consists of redundant loops rooted at distribution hubs, with short branches

hanging from towers on those loops. Every tower must appear exactly once, and reversing a cable

segment can change its cost.

How large is the search space?

There is no single closed-form count because a solution can use any number of loops, variable

loop sizes, and differently anchored and ordered branches. But Paris alone gives a useful lower

bound.

Even if we ignore ordering and branches and only assign each of the 532 terminals to one of

11 distribution hubs, there are 11^532 possible assignments.

A stronger lower bound comes from one deliberately restricted family of valid solutions:

exactly 19 loops of 28 terminals each, with no branches. This covers all 532 terminals because

19 x 28 = 532, while staying below the 30-terminal limit for a loop. Order the 532 terminals,

split that ordering into 19 consecutive groups, divide by 19! because the set of loops is

unordered, and choose one of the 11 hubs for each loop:

(532! / 19!) x 11^19 ~= 10^1223

What I tested

The primary experiment was intentionally narrow:

SettingValue

ModelsClaude Fable 5, Opus 4.8, Sonnet 5; GPT-5.6 Sol, Terra, Luna

ModesPlain; native /goal

Optimization budget30 minutes

Outer agent timeout1,900 seconds

ReasoningMaximum available setting for every model

ExecutionHarbor 0.1.43, Docker, subscription authentication

Results

Before concentrating repetitions on the flagship pair, I ran one matched 30-minute no-hint

pair for every model in the sweep. For Fable and Sol, the chart uses Pair 1 from the replicated

headline set; the other four models have one pair each.

I then repeated the flagship comparison until I had three matched runs for Fable 5 and three

for Sol.

ModelRunPlain/goalGoal minus plain

Fable 5132,19731,934-263

Fable 5232,51632,324-192

Fable 5332,44635,178+2,732

GPT-5.6 Sol133,58139,371+5,790

GPT-5.6 Sol235,53932,703-2,836

GPT-5.6 Sol333,66333,313-350

Negative means /goal was better. Goal won four of six trials, so win rate alone makes the

feature look useful. The means tell the other half:

ModelPlain mean/goal meanMean effectMedian effect

Fable 532,38633,145+759 worse-192 better

GPT-5.6 Sol34,26135,129+868 worse-350 better

Both models usually got a small benefit and occasionally suffered a large regression. That is

why /goal won most runs but made both means worse.

Fable was also clearly stronger. Its plain mean beat Sol’s by 1,875 points, and its goal mean

beat Sol’s by 1,984. More importantly, Fable plain stayed inside a tiny 319-point range while

Sol plain spanned 1,958 points. Fable goal produced the best clean score, 31,934; Fable plain

was the safest configuration.

Deep dive into the goal command

The same command hides two different systems

Claude Code and Codex both expose /goal, but the implementations are fundamentally different.

Claude Code: a separate evaluator

Claude Code implements /goal as a session-scoped Stop hook. After each main-model turn, a

small evaluator model, Haiku by default, reads the condition and conversation. It returns yes

or no with a reason. A no starts another turn; a yes clears the goal.

The evaluator cannot use tools or inspect files. It can only judge evidence that appeared in

the transcript. That can catch an early exit, but it cannot know whether another ten million

solver iterations are worthwhile. Anthropic’s goal documentation

Keep in mind that claude code is not open source, so we rely solely on what Anthropic tells us.

Codex: persisted state and lifecycle tools

I also read the source for the benchmarked release,

Codex CLI 0.144.4. Codex treats a goal as

persisted thread state:

• The TUI saves the objective for the active thread, and SQLite stores its status and budget

accounting. TUI,

schema

• The working model receives create_goal, get_goal, and update_goal tools.

Tool specification

• If the thread becomes idle while the goal is active, Codex injects a continuation turn with

the objective and a completion audit. Runtime,

prompt

Claude delegates completion to another model. Codex lets the working model declare completion,

then resumes it while the persisted goal remains active. Claude’s evaluator is independent but

sees only the transcript; Codex sees the files and tools but effectively grades its own work.

Why /goal can win most runs and still be a bad default

On a normal coding task, progress is often legible: another turn can fix a test or complete a

migration. Optimization is different. Once an agent chooses a solver, extra time can amplify

either a good decision or a bad one.

That is exactly what happened here. Goal helped when it sustained Fable’s fast compiled

portfolio or Sol’s successful chain repartition. It hurt when Fable built a slow solver or Sol

committed to an exhaustive anchor sweep. The median moved slightly in the right direction; the

bad tail moved much farther in the wrong one.

Limitations

This is one unpublished NP-hard task, not a general coding leaderboard. Only Fable and Sol have

three clean matched pairs. Other comparisons mix prompts, wrapper versions, and time limits,

and the trials ran sequentially through subscription services that may have drifted.

The containers exposed eight CPUs despite task metadata declaring one, which favored Fable’s

parallel portfolios. Every scored Fable and Sol output was valid, partly because the wrapper

required early checkpoints and final verification. The benchmark measures the complete system:

model, CLI, prompt, subscription service, and harness.

Reproducing this

The benchmark task, wrappers, analysis scripts, figure generator, and full evidence memo are in

CLIArena. Raw job directories are excluded from Git

because of their size, but the memo records every publishable score, city breakdown, elapsed

time, strategy, exclusion, and run ID.

The primary commands are:

RUN_ID=article-kiro-YYYYMMDD-clean \
PHASE=nohint-all \
./scripts/run_subscription_article_matrix.sh

uv run python scripts/summarize_subscription_article_results.py RUN_ID...
uv run python scripts/analyze_subscription_article_results.py RUN_ID...

The result I would put in the headline is not that goal helps or hurts. It is that a persistence

feature can win most individual trials while making observed average performance worse. On a hard

optimization problem, the quality of the loop matters less than the quality of what the loop

keeps doing.