What happens when you put the two most powerful AI coding agents head to head on a problem so hard that its search space exceeds 10^1223 possible configurations? Independent researcher Charles Azam ran exactly that experiment, pitting Anthropic's Claude Fable 5 against OpenAI's GPT-5.6 Sol on KIRO, an unpublished NP-hard fiber-network optimization problem from a French engineering hackathon. The results are striking: Fable 5 produced the best overall solution with a consistency that Azam describes as unlike anything he has seen from any model on this problem. But the experiment also delivered a more nuanced finding about the /goal persistence feature that should make every founder think twice before treating it as a universal improvement.
The Problem: Network Design at Astronomical Scale
KIRO is not a standard benchmark. It is a real-world fiber-network design problem Azam worked on as an engineering student in 2018, solving it in C++ over the course of a week. The task is to connect distribution points and terminals across three French cities (Grenoble, Nice, and Paris) using loops and short chains, minimizing total cable length while respecting structural constraints. The Paris instance alone has 532 terminals to assign to 11 distribution hubs. Even ignoring ordering and branches, the number of possible assignments is 11^532. One deliberately restricted family of valid solutions still produces a lower bound of approximately 10^1223 possible configurations. That is not a typo. For context, the observable universe contains roughly 10^80 atoms. This problem has more possible solutions than there are atoms in a billion universes.
Raw Intelligence Wins: Fable 5's Dominant Performance
Azam ran each model for exactly 30 minutes with maximum reasoning settings, comparing plain mode against the native /goal command. The headline result is unambiguous: Fable 5 is the stronger model for this class of optimization. Its plain-mode solutions stayed within a remarkably tight 319-point range across three runs, while GPT-5.6 Sol's plain-mode results spanned 1,958 points. Fable 5 produced the best clean score of any run at 31,934. Its plain mean of 32,386 beat Sol's plain mean of 34,261 by 1,875 points. The consistency advantage matters even more than the absolute score. For founders relying on AI agents for code generation or complex problem solving, a model that delivers repeatable quality is often more valuable than one that occasionally hits a higher peak but just as often falls into a deep valley. Fable 5 demonstrated that kind of reliability on this benchmark.
The /Goal Paradox: Winning Most Runs While Making the Average Worse
This is the finding that deserves the most attention. The /goal command, which tells the agent to keep working until a stated objective is met, won four of six individual trials. By win rate alone, it looks like a useful feature. But the means tell a different story. For Fable 5, the plain mean was 32,386 while the /goal mean was 33,145, an average degradation of 759 points. For GPT-5.6 Sol, plain mean was 34,261 versus /goal mean of 35,129, a degradation of 868 points. Both models usually got a small benefit from /goal and occasionally suffered a large regression. That is why the feature can win most runs and still make both means worse. Azam's analysis reveals the underlying mechanism. 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 strategy, extra time can amplify either a good decision or a bad one. 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 that was never going to pay off within the time budget. The median moved slightly in the right direction. The bad tail moved much farther in the wrong one.
How /goal Works Differently in Claude Code and Codex
Azam also dug into the implementation differences, which are illuminating. In Claude Code, /goal is implemented as a session-scoped stop hook. After each main-model turn, a small evaluator model (Haiku by default) reads the goal condition and the conversation transcript, returning yes or no with a reason. The evaluator cannot use tools or inspect files. It can only judge evidence that appeared in the transcript. In Codex (OpenAI's CLI), /goal is treated as persisted thread state stored in SQLite. The working model receives create_goal, get_goal, and update_goal tools it can call directly. If the thread becomes idle while the goal is active, Codex injects a continuation turn with the objective and a completion audit. The difference is fundamental: Claude delegates completion judgment to an independent evaluator that sees only the transcript, while Codex gives the working model tools to self-assess progress but effectively lets the same agent grade its own homework. Neither approach is obviously superior, but the difference means that the same /goal command can produce very different behaviors depending on which platform you are using.
What This Means for Builders
This benchmark carries several practical lessons for founders building with AI agents. First, raw model intelligence still matters. Fable 5's structured reasoning advantage produced better solutions more consistently, even without prompt engineering tricks. If your use case involves complex optimization or multi-step reasoning, the base model quality is the most important variable. Second, the /goal feature is not a generic try-harder switch. It changes the control loop and the search path, which can help or hurt depending on the problem structure. For production systems where consistency matters, you should benchmark both modes rather than assuming goal mode is always better. Third, the implementation of /goal varies significantly between platforms. If you rely on this feature for critical workflows, understand which architecture you are using and what the evaluator can and cannot observe. Finally, this kind of transparent, reproducible benchmarking is invaluable for the ecosystem. Azam published all code, prompts, result tables, exclusions, and trajectory notes in the CLIArena repository. Every founder building AI-powered products should bookmark this methodology and consider running similar controlled experiments on their own use cases. The models change fast, but the engineering discipline of measuring what actually works never goes out of style.
