OpenAI Introduces o3, its most capable reasoning model to date

OpenAI’s “12 Days of Surprises” advent calendar, from the creators of ChatGPT, has unveiled o3, the most advanced prototype in its o1 reasoning model series. This AI also comes with a mini version, known as o3-mini, which is a faster, more streamlined model optimized for coding.

While these AI models are not yet available to the general public, OpenAI is offering access to security researchers who apply for testing to help in their evaluation process. As for the launch, the company plans to release o3-mini in early 2025, with o3 following soon after.

Sam Altman, CEO of OpenAI, announced this significant update and explained why the second release in their advanced reasoning AI series is named o3 instead of 02. This decision was made to prevent any mix-up with the telecommunications provider O2.

Today, we shared evals for an early version of the next model in our o-model reasoning series: OpenAI o3 pic.twitter.com/e4dQWdLbAD

— OpenAI (@OpenAI) December 20, 2024

Exceptional software and coding abilities

o3’s reasoning skills exceed those of its predecessor, enabling it to tackle more complex problems. This means that it will take slightly longer to formulate responses compared to models in the GPT series, although its reasoning duration is adjustable. Unlike o1, o3 can be set to low, medium, or high computation times, depending on the task’s complexity.

During the introduction of this innovation, o3’s capabilities and its impressive performance on technical standards were highlighted. For instance, in coding benchmark tests, it was demonstrated that o3 achieves 71.7% accuracy in software tasks, which is over 20% higher than the 48.9% accuracy of o1.

In terms of coding tasks, o3 scores 2,727 on Codeforces, a coding proficiency benchmark, while o1 scored 1,891. During the presentation, Mark, an OpenAI researcher alongside Altman, mentioned that his highest score as a professional programmer in a similar test was 2,500.

OpenAI has even claimed that o3 is a significant step towards general artificial intelligence. This assertion seems overly ambitious to us, as while this model does represent a leap forward in reasoning abilities and executing complex tasks, it doesn’t fully embody the traits that define AGI.

The company previously described AGI in December 2023 as follows: “General artificial intelligence (AGI) is the capacity of a machine to perform intellectual tasks that typically require human intelligence. Unlike specialized AI, which focuses on specific tasks, AGI aims to replicate human cognitive versatility across a broad spectrum of activities.”

Other Holiday Updates from OpenAI

This marks the conclusion of OpenAI’s 12 days of surprises, which they organized as a holiday gift for their audience. Throughout this event, several major features were announced, such as the real-time vision, the introduction of SearchGPT for all ChatGPT users, and the ability to communicate with ChatGPT via WhatsApp.

Additionally, as a special bonus, Sam Altman announced that all Plus plan subscribers will have unlimited access to Sora, an AI video generator, via a “reduced wait queue” over the Christmas holidays. This means they can use the tool more efficiently and faster.

Day 13 of Shipmas: special Sora bonus✨

Our GPUs have less activity in late December as people take time off from work, so we’re giving all Plus users unlimited Sora access through a reduced wait queue during the holidays!

Enjoy creating!

— Sam Altman (@sama) December 21, 2024

Photo: Flux Schnell