OpenAI’s 12 Days of OpenAI campaign wrapped up today with two major announcements: they unveiled o3 and o3 Mini models. Just days after the full release of the o1 model, OpenAI has outdone itself by announcing an even better-performing reasoning model. Here’s everything you need to know.
Table of Contents
1. o3 Reasoning Model
OpenAI’s o series AI models are reasoning models, meaning they take time to think step-by-step to arrive at a conclusion. This approach enables better accuracy and allows these models to solve complex problems, particularly in programming, math, and science.
The full version of the o1 model was released on Day 1 of the “12 Days of OpenAI” campaign. However, as competition intensifies with models like Gemini 2.0 Flash Thinking, OpenAI has raised the bar with the o3 model, which demonstrates significant improvements over the o1 model in both benchmarks and accuracy.
For example, in programming tasks, the model’s accuracy improved from 48.9% to 71.7%.
The o3 model also showed notable advancements in solving math problems and PhD-level science questions.
Additionally, it achieved a significant improvement in the ARC Prize benchmark (Abstraction and Reasoning Corpus for Artificial General Intelligence). Basically, it tests an AI’s ability to solve unfamiliar puzzles using logic and pattern recognition. Instead of relying on previous training or data, the benchmark focuses on how well the AI can think and generalize to solve new problems, similar to human reasoning.
2. o3 Mini: Cost-Efficient Model
Along with the o3 model, OpenAI also announced o3 Mini, a model designed for speed and efficiency.
o3 Mini focuses on low, medium, and high-reasoning tasks, offering a balance of capability and cost-effectiveness. In programming benchmarks, while o3 Mini (Low) achieved the same Elo rating as the o1 Mini, the o3 Mini (High) outperformed even the full o1 model.
A similar trend was observed in math benchmarks. However, it’s worth noting that while o3 Mini (Low) having the same accuracy of the o1 Mini, it achieves this with greater efficiency, reducing the latency of its outputs.
During the live demo, o3 Mini demonstrated its versatility by writing scripts that evaluated its own performance.
Also Read:
- OpenAI Introduces Reinforced Fine-Tuning for o1-Mini Model
- ChatGPT Can Now See and Talk to You: Advanced Voice Mode with Video Launched
- ChatGPT Expands “Work with Apps”: More Apps, Search, and Voice Integration
Availability
Both o3 and o3 Mini models are not yet available to the public. OpenAI is prioritizing safety, granting early access exclusively to select researchers for external safety testing. Applications for external safety testing opened today, and you can apply from here.
While a specific release date was not provided, o3 Mini is expected to launch shortly after the release of o3.
The End of 12 Days of OpenAI
OpenAI’s campaign concludes today, and among all the announcements, Sora, its text-to-video generation model, stands out as the most exciting. Updates to Canvas and Search are also highly practical and beneficial for a broad range of users.
While the updates to the reasoning models, such as o3 and o3 Mini, represent a significant leap in AI capabilities, they may be less relevant for users who are not into solving PhD-level problems.
