GPT-5 for one-shot simulation

A week into using GPT-5 feels like a good moment to step back and share some early thoughts. I’ve been trying to avoid reading too deeply into other’s opinions of the model, though that’s been pretty hard to do on X this past few days. My own take is bound to be coloured by my experience, preferences and approach to using other models; but with the twitterverse seemingly polarised by GPT-5, I’ve been eager to give it an impartial hearing.

So, first impressions: GPT-5 clearly delivers a meaningful jump in raw intelligence. It's not the paradigm shift we might've been dreaming about, but still a significant upgrade, particularly for power users. For more casual users experiencing reasoning models for the first time, this is bound to feel like a bigger deal. How big a step-up really depends on your workflow, use case, and personality preferences.

Coding is a key use case for me, and so far GPT-5 has been impressive. I've not yet taken it for a proper spin in Cursor - I’ll save that fun for this weekend - but straight out of the box and into Canvas, it’s shown real promise.

Here's the prompt I threw at it:

create a js app comprising a rotating circle with a 30 degree section removed. a coloured ball is ejected from the centre at a random angle and falls under gravity, bouncing on the inside edge of the circle. a small amount of friction on the inside edge of the circle changes the trajectory of the ball. when it exits through the gap in the circle, two more balls appear at the centre. the balls can collide with each other. run the simulation till the user hits stop.

You can see the results below. GPT-5 thought that over for 1 minute and 35 seconds, then produced an impressive one-shot result; just under 500 lines of code, generating a polished UI with eight physics sliders, clear user controls, and informative read-outs. For comparison, I ran the same prompt through o3 and Opus 4.1, both of which also nailed a simulation in one go, but with nothing like GPT-5's refinements and attention to detail. That said, Claude's retro-inspired interface has its own charm - UI aesthetics are all about Taste after all!

One thing I couldn’t avoid hearing about GPT-5 is that it responded poorly to vague prompting, preferring the more rigid, CoT style approach necessary to get the most from 3 and early 4 series models. Honestly, I haven’t encountered this yet. I'm as guilty as anyone of leaning into the freedom reasoning models have given us from structured, bulky JSON prompts, and wasn't keen on going back. Thankfully, even when my prompts have been characteristically loose, side-by-side tests suggest GPT-5 consistently matches or improves on outputs from legacy models.

So far, so good. I’ll keep pushing it to see where else GPT-5 shines or stumbles, but I already like what I’m seeing in the coding space.