GTA V in an Afternoon?

My last post got a surprising amount of engagement after I made the tongue-in-cheek claim that Claude Fable could generate GTA V in an afternoon.

Obviously, it can’t. Anyone who knows anything about game development knows that.

Still, the joke made me wonder how close we could get.

So I opened Fable, set it to high effort and gave it this painfully specific prompt:

“Please create me GTA V.

Make it one single HTML file. Loop and test until it’s perfect.

Make no mistakes.”

One prompt and 11% of my premium subscription session limit later, I had an extremely playable version of Grand Theft Auto.

It was much closer to GTA III than GTA V, but it had carjackings, police chases and enough of an actual game loop to keep me playing. All of it was contained in one HTML file.

Play Gran Theft Otto

Since that had barely touched my time or session limits, I sent it one more prompt:

“This is great, but I was hoping for a more first person experience.

Do it please. And make no mistakes.”

That took me to 26% of my session allowance and produced another working game.

Play Gran Theft Otto: First Person

ChatGPT has a go

These had been so quick to make that deploying the HTML files had taken me longer than generating them.

Naturally, I decided to see what ChatGPT could do.

I gave the same job to GPT-5.6 Sol on high effort, using an entry-level subscription. Sol split its path and gave me two first iterations to choose from in the usual “Which answer do you prefer?” style.

Both were better than the Fable output.

I really wasn’t expecting that, especially given the difference in subscription price. It has definitely made me more likely to use the OpenAI models in my day-to-day work.

The first option was Metro Vice:

Play Metro Vice

The second was Metro Mayhem:

Play Metro Mayhem

Metro Mayhem was genuinely good

Metro Vice was decent, but Metro Mayhem was easily the best result of the experiment.

It took much longer to finish, but the end result was genuinely impressive. I’ve already wasted more time than I should driving around and attempting the missions.

They are harder than you might expect. The NPCs are aggressive, there are no map markers and I am apparently terrible at driving a pretend car.

Since Metro Mayhem was so far ahead of the others, I gave Sol the same first-person prompt I’d sent to Fable:

While I waited, I tried DeepSeek.

Thirty seconds with DeepSeek

I’d never really used DeepSeek before. Most of my AI work requires me to opt out of having my data used for training, so it hasn’t usually been an option.

This experiment was hardly commercially sensitive, though, so I gave it the original prompt.

Within about 30 seconds, DeepSeek produced something that felt like ants chasing each other around a square.

It wasn’t remotely close to the depth of the Fable or Sol games, but I was using a free account. It seemed unfair to complain too much.

So I did what any reasonable person would do and asked it to make the game 3D.

Thirty seconds later, it did.

Play DeepSeek’s GTA Mini City

Play GTA Mini City 3D

Make no mistake: the free DeepSeek output is rubbish. But it is instant and it is free, which makes it difficult to dismiss completely.

I’d be interested to see what its paid frontier models could produce. Maybe one day.

Back to Metro Mayhem

After my brief DeepSeek detour, I went back to ChatGPT to see how it was getting on with the first-person version of Metro Mayhem.

Unfortunately, this was where the winning streak ended.

The result ate through my browser’s memory and wasn’t much fun to play. Given how good the original Metro Mayhem was, it was easily the most disappointing result of the lot.

Play Metro Mayhem: First Person

Are software engineers finished?

No.

These experiments produced some surprisingly playable games, but they also produced broken controls, bizarre NPC behaviour, missing map markers, performance problems and, in DeepSeek’s case, ants chasing each other around a square.

What they did prove is that turning instructions into code is no longer much of a moat for software engineers, assuming it ever was.

The models can get surprisingly far on their own.

They still need someone who knows when the result is rubbish.