OpenAI accidentally leaked its o1 model, and it looks incredible

Last week, during a Reddit Ask Me Anything (AMA), OpenAI CEO Sam Altman previewed OpenAI’s full o1 model as a “noticeable improvement” over the o1-preview we currently have access to.

Those claims were put to the test for about two hours on Friday, as OpenAI accidentally leaked the full o1 model.

What is o1?

o1 — previously codenamed "Strawberry" — is the latest series of AI models from OpenAI. The main difference between o1 and OpenAI’s other GPT models is that o1 focuses on inference, while GPT focuses on data. 

In other words, o1 gets better the longer it can “think” about a problem, while GPT improves as it trains on more data.

So, because o1 is better at “thinking,” it is especially good at stuff like translating a corrupted sentence, coding a video game from a prompt, and, for the first time in the history of AI, answering how many r’s are in “strawberry.”

The leak

OpenAI has so far only released o1-preview, which is a smaller version of the full o1 model that merely previews its vast potential. 

This is why when people were so excited when they were able to access the full model by just changing a parameter in the URL.

Unfortunately, the leak only lasted two hours, so it wasn’t possible to really play around with it. But, from the small sneak peek we did get, the model looks to be especially good at interpreting images. 

For example, here’s it solving a fairly complex image puzzle.

And here’s it going into very specific detail on an image of a SpaceX rocket.

Of course, it isn’t just better at interpreting images. It also appears to have received a pretty substantial reasoning and coding buff. 

Overall, it looks like Altman wasn't lying when he called it a "noticeable improvement."

When is it coming out for real?

At this point, only OpenAI knows. 

What we do know from the Reddit AMA is that OpenAI has a substantial release coming later this year. Considering that Altman says it isn’t “something we are going to call GPT-5”, then the most likely candidates appear to be either Orion (which is basically GPT-5 under a different name) or the full o1 model.

In any case, it seems like we’ll be receiving more AI fireworks before year's end.