OpenAI Unveils Next-Gen o3 Model Set for Early Next Year Release

OpenAI has previewed its next-generation frontier model, o3, promising improved performance and new features, with a mini version launching first.

OpenAI just wrapped up its ’12 Days of OpenAI’ livestream with a bang, unveiling the next big thing: the o3 model (and its little sibling, o3-mini). Named with a cheeky nod to the O2 network and OpenAI’s own quirky naming habits, o3 isn’t ready for your average Joe yet. First up, it’s going to researchers—because safety first, right?

Here’s the scoop: o3-mini is hitting the shelves by January’s end, with o3 hot on its heels. And let me tell you, o3 isn’t just a step up from o1; it’s a giant leap. Crunching numbers? It aced the American Invitational Mathematics Examination with a 96.7% accuracy, leaving o1’s 83.3% in the dust. Not too shabby, huh?

But wait, there’s more. OpenAI’s thrown down the gauntlet with tougher benchmarks like ARC-AGI, really putting o3 through its paces to see if it can think on its feet. And think it did—scoring 75.7% on low power and a whopping 87.5% when let off the leash, breezing past the human benchmark of 85%. Talk about overachieving.

As for o3-mini, it’s packing the Adaptive Thinking Time API, letting you tweak how long it spends noodling over problems (Low, Medium, or High). It’s like o1, but cheaper and just as smart. So, whether you’re pinching pennies or just need something that gets the job done, o3-mini’s got your back.