I Made GPT-5.6, Claude Fable 5, and Grok 4.5 Build a Football Game. The Cheapest One Won.

You’ve probably seen the headlines. GPT-5.6 Sol is the “smartest model ever.” Claude Fable 5 is a “reasoning powerhouse.” Grok 4.5 is… well, Elon’s latest toy.

I decided to stop reading benchmarks and start breaking things. I made all three models build the same thing from scratch: a 5-a-side football game. No templates. No human intervention. Just pure, unfiltered AI coding.

What happened next was a beautiful disaster. The $150/mo model produced a game that felt like a fever dream. The $10/mo model actually worked.

Let me take you through the carnage.

The Setup: A Fair Fight to the Death

I used Cursor’s sub-agent feature to launch all three models simultaneously. Same prompt. Same environment. Same 10 functional requirements: physics engine, AI opponent, match system, leaderboard, stats panel, the works. A football game is the perfect stress test because it demands real-time physics, state machine logic, and canvas rendering. You can’t fake it. The result is either a playable game or a glitchy mess.

I told them to code, test, find bugs, and fix them in a loop. Zero human oversight.

The Battle Begins

Claude Fable 5 immediately did what Claude does: it spent 5 minutes writing a detailed plan. Modules, architecture, dependencies. It was the academic who outlines the essay before writing the first sentence.

GPT-5.6 Sol and Grok 4.5 just started typing. They wrote code, hit run, saw it break, and fixed it. They were the scrappy startups moving fast and breaking things.

Speed difference was brutal. GPT finished in 7 minutes with 2 bug-fix cycles. Grok took 11 minutes. Claude took 20 minutes. In a rapid iteration scenario, planning is a liability.

But speed means nothing if the output is garbage. So I played the games.

GPT-5.6 Sol: The Speedy But Hollow Winner

First impression: solid. Clean UI, no login wall, you’re straight into team selection. My team was called “Neon Dragon” which was either awesome or cringe depending on how drunk I was.

Then the match started. I stood still for 20 seconds to see the AI opponent’s strategy. The enemy striker just headbutted my goalkeeper repeatedly. Neither scored. It was like watching two drunk friends argue about who has to pay the bill. Eventually, my own players started fighting each other. Literally. Two red teammates tackled each other and the ball just sat there, abandoned.

I played an entire match on medium difficulty. Final score: 0-0. The AI defense was impenetrable because it was broken. On easy mode, my striker “Li Football” scored a solo goal. Victory. But it felt empty.

GPT-5.6 Sol delivered a technically complete game that was functionally boring. It passed all 10 requirements. It was just… not fun.

Mobile view? Absolute disaster. The pitch was squished into a postage stamp.

Grok 4.5: The Ugly Underdog That Actually Tried

Grok’s login screen was a psychedelic nightmare. Colors that didn’t belong in a sports game. My team was called “Galactic Battleship” and the opponent was “Star Alliance.” At least it had a theme.

The actual gameplay was hilariously bad. My goalkeeper was clearly a statue in a previous life. He watched the ball roll past him. Three times. Same opponent scored all four goals. My teammates suffered from collective narcolepsy, leaving the goalie to duel alone.

Final score: 4-0. I lost. But I was laughing. The code was modular and clean (8 files, 2622 lines), but the game was broken. Grok’s swing was wild, but it swung hard.

Mobile view actually worked. The pitch was visible. The controls were usable. It was the most functional of the three on mobile, which is ironic given how bad the desktop experience was.

Claude Fable 5: The Most Expensive Disaster

This is where it gets painful. Claude’s login page was fine. Generic. Safe. The team selection screen used emojis so aggressively it felt like a children’s coding tutorial.

Then the match started. The players looked like… chickens. I’m not joking. The sprites were tiny, featureless, and vaguely bird-like. But the real horror was the ball. It teleported. Every two seconds. The physics engine was broken on a fundamental level. I watched the ball vanish from a player’s foot and reappear behind the goal. The AI opponent had mastered the art of interdimensional football.

I quit after one minute. It was unplayable. Not bad. Unplayable.

Claude took 20 minutes, used massive thinking mode, produced 1844 lines of code, and delivered a game that defied the laws of physics. The mobile view was actually the best of the three, but that’s like saying a car with no engine has the best paint job.

Claude Fable 5 failed because it overthought. The extended reasoning mode spent so much time on architecture that it forgot to make the ball stay in one place.

Code Quality: The Numbers Don’t Lie

All three models built all 10 features. But how they built them tells the real story.

GPT-5.6 Sol: 683 lines, single file, 2 bug fixes. Efficient to the point of being reckless. Code density was high, but it worked.

Grok 4.5: 2622 lines, 8 files, modular architecture. Cleanest codebase. The game was broken, but the code was beautiful. A paradox.

Claude Fable 5: 1844 lines, but the core game logic was in a single 33KB file. A maintenance nightmare. The architecture was overengineered in the wrong places.

All three used hand-written physics engines. None used a library. They all implemented rule-based state machines for AI opponents. Same approach, wildly different results. The gap between code logic and running effect is still massive.

The Final Ranking

1. GPT-5.6 Sol: Won by default. It was the only one that produced a playable game. Not a good game. A playable one. But that’s the bar.

2. Grok 4.5: Second place. The game was comically broken, but it was fun to break. The code quality was the best. It just didn’t translate to a good experience.

3. Claude Fable 5: Last place. The most expensive model produced the worst result. The ball teleporting bug is a fundamental physics failure. Ten minutes of “thinking” didn’t help.

This is the part that should terrify and delight you simultaneously.

The most expensive model isn’t always the smartest. The longest thinking time doesn’t guarantee the best output. And sometimes, the cheapest model wins by failing faster and iterating harder.

What This Means For You

If you’re building a quick prototype, a small game, or any project that requires rapid visual feedback, pick GPT-5.6 Sol. It’s half the price of Claude Fable 5 and it ships faster.

If you need complex architecture and long-running tasks, Claude Fable 5 is still your best bet. But don’t ask it to make a game. That’s like asking a philosopher to play dodgeball. They’ll overthink getting hit.

If you’re on a budget, Grok 4.5 is a legitimate option. It’s 5x cheaper than Sol and 10x cheaper than Fable 5. The code quality is solid. The execution is just inconsistent.

But here’s the uncomfortable truth: none of these models produced a genuinely good game. Not one. They all failed at basic gameplay intelligence. The AI opponents didn’t coordinate. The physics were janky. The fun was missing.

Human developers are still safe. For now. The AI can write the code, but it can’t make it fun. And in game development, fun is the only metric that matters.

The benchmark scores are a lie. The price tags are a suggestion. The real test is picking the right tool for the job. And sometimes, the right tool is the one that costs less and breaks faster.