OpenAI is expected to release a new flagship model called GPT-5 sometime this year or early next year, although Sam Altman has said that he's not sure if they'll actually call it GPT-5.
This market will resolve based on the official, confirmed public release of that anticipated model. If it is not named GPT-5 then Altman/OpenAI must make it clear that it is the same anticipated model which everyone has been expecting to be called GPT-5. If this is not clear, a model not named GPT-5 will not count for this resolution.
A test release like the GPT2 chatbots put up on Chatbot Arena before the release of GPT-4o would not count for this market, and neither would something like the preview release of GPT-4 through Microsoft Azure. This market requires a broad, official release which is either free or available at a reasonable cost like through a GPT plus membership.
These exact resolution criteria may be updated to better match the spirit of the question, and your suggestions are welcome in the comments.
All options in this market resolve NO as that month arrives until the anticipated model is broadly released and all remaining options then resolve YES.
1,000This doesn't have anything to do with resolution, but I find it funny that the model that was expected to be called GPT-5 (by the people who created it) was already released, but they called it GPT-4.5 because it wasn't impressive enough.
By November, I was pretty strongly expecting the Orion model to be released in February, and I nailed it. But they didn't put the GPT-5 brand name on it in the end, so I lost out. 🥲
@jim I was just being chatty, but since you asked so nicely, I went and found one. 😄
https://www.wsj.com/tech/ai/openai-gpt5-orion-delays-639e7693
Edit: two, to connect 4.5 to Orion as well.
https://techcrunch.com/2025/02/27/openai-unveils-gpt-4-5-orion-its-largest-ai-model-yet/
@Haiku But is GPT-4.5 100x the training compute/FLOPs of GPT-4? My impression was that it's more likely only 10x, and so 4.5 is quite well deserved. And GPT-4o is not 4.5, because it's only ~2x (I think slightly less, but I can't remember the exact estimate from EpochAI).
Of course the actual amounts for 4.5 are unconfirmed, but just in terms of what's most likely the correct ballpark I think 10 >>> 100 (but am open to changing my mind based on evidence)
oh never mind, the guy has a relevant calculation somewhere I think, but I don't have the time to search for it
@adssx I hope you know that if the original tweeter had instead asked, “Weeks? Months? Years?”, then Altman would have answered, “weeks / months / years”.
@NBAP The slash in SamA's tweet refers to the slash in Chubby's tweet. He didn't mean they would both be released in a matter of weeks or months, but that GPT-4.5 would be release in a matter of weeks and GPT-5 would be released in a matter of months.
@Haiku Yeah, it’s possible that’s what he meant, given that it seems to actually be the case. But it would also be entirely in character for him to just be shitposting.
It’s hard for the neural networks to be open-source (I'm sure OpenAI themselves probably doesn’t even keep their own “source code” lest they get caught red-handed with a provenance trail for scraped YouTube videos and pirated text bodies); I wasn’t implying anything about that.
I was referring to release of the models, actually physically publishing the weights and biases — releasing the model, as opposed to an API that provides limited access to the model.