GPT 4.5 - the future ain't what it used to be

Once upon a time the only way to make a language model smarter was to train it on more data.

This led to a belief that scale was everything and that only the tech giants with their access to an internet’s worth of training data and millions of GPUs would be able to create AI.

Then reasoning models came along - the arguably most famous of which, Deepseek, upended this expectation (at least if you weren’t paying attention) as it became clear that you could get very capable AI at a much lower cost.

But this left a problem for OpenAI. The old way of training models takes many months (according to some rumours GPT-4.5 first started being trained back in Feb 2024). And by the time they completed the process they discovered that these new reasoning models outperformed their old approach.

Clearly they couldn’t claim their new model was GPT-5 as it wasn’t as capable as the reasoning models built off of GPT-4. So they decided to claim they hadn’t really been trying all that hard and actually this was just an iteration - half a model if you will, which is how we end up with GPT-4.5.

The interesting thing with GPT-4.5 is that it shows that there is still milage left in the original scaling approach - it’s much better than the previous model in it’s class, GPT-4o. And it’s also the case that new reasoning models built on top of it should way out-perform the current generation.

But the truth is that OpenAI have become a victim of their own success - they were the first to introduce reasoning models and now those reasoning models are making the old kind less relevant.

That almost certainly means that GPT-4.5 will be the last non-reasoning model we’ll see.

Yes, there are plenty of use cases where reasoning doesn’t seem to be all that helpful (eg creative writing) but I’d hazard a bet that’s because we haven’t quite got the reasoning right for those types of tasks rather than that it won’t work.

In some ways this is good as it means some of the potential barriers to developing AI (such as using up all the world’s training data) are no longer such a concern but reasoning models are far more energy intensive than standard models and therefore their ascendancy will cause further questions over AI’s environmental impact.

If there’s one key lesson to draw from this moment is that even the leaders of the pack can’t predict 12 months ahead. OpenAI invested a huge amount in an approach that a new algorithm made much less relevant. If they can’t get it right what hope do the rest of us have?

To that end it remains key for every business to think about how they can make themselves as flexible and adaptable as possible so that when the AI driven economic shockwaves hit they’re as well placed to survive as possible.