OpenAI released GPT-4.5 this week to Pro subscribers and via the API, but it’s not what I was expecting. It is much more expensive, about 15-30 times the cost of GPT-4o. For my simple needs, like figuring out the keywords in a photo or summarizing a web page, older models are fine.
Sam Altman posts on Twitter / X (sigh) that it’s a giant model. OpenAI can’t roll it out to Plus customers until they bring online tens of thousands of new GPUs. Sam adds:
this isn’t a reasoning model and won’t crush benchmarks. it’s a different kind of intelligence and there’s a magic to it i haven’t felt before.
Simon Willison suggests that 4.5’s training knowledge cut-off is not any different than 4o, only going up to the end of 2023. This makes me wonder if data crawling and licensing issues have derailed the company to an extent, and they’re holding the new data for GPT-5. They say that improvements for 4.5 were partially from synthetic data created by other models.
There is also this from OpenAI’s blog post:
GPT‑4.5 is a very large and compute-intensive model, making it more expensive than and not a replacement for GPT‑4o. Because of this, we’re evaluating whether to continue serving it in the API long-term as we balance supporting current capabilities with building future models. We look forward to learning more about its strengths, capabilities, and potential applications in real-world settings.
Translation: we don’t really know how this model works. 🤪
I remain fascinated with AI. There are many people who are worried, and if that’s you there’s not much new here that will reassure you, except some good news with a lower hallucination rate. I’ve also blogged myself about the potential harm of agents in particular.
But AI might be the last truly new thing I’ll see in my programming career. We are so used to exciting new gadgets and software, released all the time, and yet none of it is profoundly new. AI is the only thing that’s comparable in scale to the tech advances that have changed everything, like the graphical user interface, mobile phones, and the web itself.
It reminds me of an old interview with Steve Jobs, on visiting Xerox PARC in 1979:
And they showed me really three things. But I was so blinded by the first one, that I didn’t even really see the other two. One of the things they showed me was object-oriented programming. They showed me that. But I didn’t even see that. The other thing they showed me was really a networked computer system. They had over a hundred Alto computers, all networked, using email, etc, etc. I didn’t even see that.
I was so blinded by the first thing they showed me, which was the graphical user interface. I thought it was the best thing I had ever seen in my life. Now remember, it was very flawed. What we saw was incomplete. They’d done a bunch of things wrong, but we didn’t know that at the time. Still though, they had… The germ of the idea was there and they’d done it very well.
And within, you know, ten minutes, it was obvious to me that all computers would work like this someday.
Perhaps GPT-4.5 represents the end of the breakthrough in generative AI, as we enter a more iterative period of refinement, the same way that the graphical user interface and object-oriented programming haven’t fundamentally changed since Steve Jobs saw them. Those things from Xerox PARC — from Smalltalk to the Alto’s mouse — are all recognizable as early versions of what we have today.
So if newer AI models are only marginally better, or only different in ways that we can feel but not measure, that’s okay. Models need to be safer and more efficient, for both developers and the environmental impact. The pace with AI over the last year has been almost too much. But the change did happen, it’s blindingly obvious, and everything is going to be a little different from now on.
