As an analogy, think of it like this:
o3-low ~ Ford Mustang with the accelerator gently pressed
o3-medium ~ Ford Mustang with the accelerator pressed
o3-high ~ Ford Mustang with the accelerator heavily pressed
o3 pro ~ Ford Mustang GT
Even though a Mustang GT is a different car than a Mustang, you don’t give it a totally different name (eg Palomino). The similarity in name signals it has a lot of the same characteristics but a souped up engine. Same for o3 pro.
Fun fact: before GPT-4, we had a unified naming scheme for models that went {modality}-{size}-{version}, which resulted in names like text-davinci-002. We considered launching GPT-4 as something like text-earhart-001, but since everyone was calling it GPT-4 anyway, we abandoned that system to use the name GPT-4 that everyone had already latched onto. Kind of funny how our original unified naming scheme made room for 999 versions, but we didn't make it past 3.
> As o3-pro uses the same underlying model as o3, full safety details can be found in the o3 system card.
- o3 pro is based on o3
- o3 pro uses the same underlying model as o3
- o3 pro is similar to o3, but is a distinct thing that's smarter and slower
- o3 pro is not o3 with longer reasoning
In my analogy, o3 pro vs o3 is more than just an input parameter (e.g., not just the accelerator input) but less than a full difference in model (e.g., Ford Mustang vs F150). It's in between, kind of like car trim with the same body but a stronger engine. Imperfect analogy, and I apologize if this doesn't feel like it adds any clarity. At the end of the day, it doesn't really matter how it works - what matters is if people find it worth using.
Here's the current state with version numbers as far as I can piece it together (using my best guess at naming of each component of the version identifier. Might be totally wrong tho):
1) prefix (optional): "gpt-", "chatgpt-"
2) family (required): o1, o3, o4, 4o, 3.5, 4, 4.1, 4.5,
3) quality? (optional): "nano", "mini", "pro", "turbo"
4) type (optional): "audio", "search"
5) lifecycle (optional): "preview", "latest"
6) date (optional): 2025-04-14, 2024-05-13, 1106, 0613, 0125, etc (I assume the last ones are a date without a year for 2024?)
7) size (optional): "16k"
Some final combinations of these version number components are as small as 1 ("o3") or as large as 6 ("gpt-4o-mini-search-preview-2024-12-17").
Given this mess, I can't blame people assuming that the "best" model is the one with the "biggest" number, which would rank the model families as: 4.5 (best) > 4.1 > 4 > 4o > o4 > 3.5 > o3 > o1 (worst).