Branding aside, this pretty much is GPT 5.

The evidence for that is the change in the tokenizer. The only way to implement that is to re-train the entire base model from scratch. This implies that GPT 4o is not a fine-tuning of GPT 4. It's a new model, with a new tokenizer, new input and output token types, etc...

They could have called it GPT-5 and everyone would have believed them.

I’ve used it for a couple of hours to help with coding and it feels very similar to gpt4: still makes erroneous and inconsistent suggestions. Not calling it 4.5 was the right call. It is much faster though.

The expectations for gpt5 are sky high. I think we will see a similar jump as 3.5 -> 4.

That doesn't imply that it's GPT-5. A GPT-4 training run probably doesn't take them that long now they've acquired so many GPUs for training GPT-5.

Pretty sure they said they would not release GPT-5 on Monday. So it's something else still. And I don't see any sort of jump big enough to label it as 5.

I assume GPT-5 has to be a heavier, more expensive and slower model initially.

GPT-4o is like an optimisation of GPT-4.

I think 4o is actually noticeably smarter than 4, after having tried it a tiny bit on the playground.

There has been speculation that this is the same mystery model floating around on lmsys chat bot arena and they claim a real observable jump on elo scores but this remains to be seen some people don't think its even as capable as GPT4-Turbo so tbd

It's quite different from gpt4 in two respects:

1) It's natively multi-modal in a way I don't think gpt4 was.

2) It's at least twice as efficient in terms of compute. Maybe 3 times more efficient, considering the increase in performance.

Combined, those point towards some major breakthroughs having gone into the model. If the quality of the output hasn't gone up THAT much, it's probably because the technological innovations mostly were leveraged (for this version) to reduce costs rather than capabilities.

My guess is that we should expect them to leverage the 2x-3x boost in efficiency in a model that is at least as large as GTP4 relatively soon, probably this year unless OpenAI has safety concerns or something, and keeps it internal-only.

It's a completely new model trained from scratch that they've decided to label that way as part of their marketing strategy.