Fable 5.2 built Theo's game in 76 minutes for $42

Theo ran his Fishslop game task on Fable 5.2 and says the model finished in 76 minutes and cost about $42 in tokens. He calls it one of the best Fishslop runs he has seen, then immediately qualifies it. In his words it was "not significantly better than Opus 5.5 (esp. for the price)".

Theo - t3.gg

@theo

X

One of the best Fishslop runs I've seen for sure, but not significantly better than Opus 5.5 (esp. for the price)

The comparison is the useful part. Theo posted a reminder of what the Opus 5.5 version of the same task looked like. That build skipped the "open tank" thing, which he says hurt it graphically, but he rates it higher on the things that make a game a game. Overall it feels more playable to him, like it got the gameplay loop better, and the controls and performance are better too.

Graphics are not the score

That split is worth sitting with. A model output that looks more impressive in a screenshot lost on the axes Theo actually cares about, and the one that looked worse won on feel. If you are judging agent output on a demo clip rather than on using the thing, you are grading the easiest part to fake.

Theo's read is that the two models are complements rather than rivals for this kind of work. He says they will work great together for 3d game type tasks, without spelling out what that split would look like in practice.

How much this proves

One run, one person, one self-designed task, with no repeated trials and no shared scoring rubric. Fishslop is Theo's own recurring prompt, so the comparison is only as good as his memory of previous runs, and "feels more playable" is a judgement call, not a measurement.

It still beats most model takes going around this week, because it is the same task on two models with a wall clock number and a dollar number attached. For working engineers, that is the shape of comparison worth copying. Pick a task you run often, note the minutes and the token spend, then say which output you would actually ship.