Ask HN: How did you like GPT-4.5?

1 points by felipemesquita 284 days ago | 2 comments

I found it was the best model when the task at hand required only “writing” - no coding, tool use or reasoning. Like proofreading and suggesting better phrasing.

I think it would fare quite well against GPT-5 in writing tasks like the ones progress.openai.com.

I’m disappointed it’s been deprecated and there where not any comparisons with it in the GPT-5 demos.

ps: if someone here is on the pro plan and that still gives access to 4.5, could you please give it the prompts in progress.openai.com to see how it compares with 5?

A_D_E_P_T 284 days ago |

GPT-4.5 was good, and I really liked it, but it's not as capable as o3/o3-Pro, and it was actually a worse writer -- i.e., a worse prose stylist -- than the Chinese models Deepseek R1 and Kimi! (Deepseek, for whatever reason, has always been incredibly good at writing in English, and has an impressively neutral voice.)

So for the prompt: "Write a poem to explain Newton's laws of physics"

GPT 4.5:

When thinking how the world does spin, / Newton’s laws tell us how things begin. / Three simple rules describe it all, / From mighty planets to a bouncing ball.

First Law, simple and quite clear: / An object moves or stays right here. / Until a push, a pull, a shove, / It stays at rest, or stays in move.

Second Law, connects three parts: / Mass and force, where motion starts. / The greater force upon the mass, / The quicker speed will come to pass.

Third Law is action, reaction: / Equal force, opposing traction. / Push a wall, it pushes back, / Force meets force, on equal track.

So these three laws, by Newton taught, / Explain the moves that nature’s wrought. / From smallest grain to starry skies, / His laws make clear how motion flies.

This is okay but very flat and pedestrian.

Kimi does a lot better, is a lot more imaginative, but gets meter a bit wrong:

> https://www.kimi.com/share/d2afi1ds8fb5aditaubg

minimaxir 284 days ago |

GPT-4.5 was a dud release which is why they followed up with GPT-4.1 and ignored 4.5's existence completely.