**OpenAI’s Christmas Countdown Live Broadcast Unveils o3 and o3-mini Models**
Starting from December 4, U.S. time, OpenAI held a continuous 12-day Christmas countdown live broadcast, first releasing the official version of model o1 on the first day. On the final day, they surprised the audience with a major announcement: the launch of the next-generation models o3 and o3-mini.
### Breakdown of the Three Highlights in ChatGPT’s Christmas Gift
Similar to o1, OpenAI’s latest models o3 and o3-mini also use the “Chain of Thought” (CoT) reasoning technology. However, their performance has been significantly improved compared to o1.
The performance of the o3 model has shown significant improvements over o1.
Image / OpenAI YouTube
During the live broadcast on December 20 (U.S. time), OpenAI’s Vice President of Research, Mark Chen, demonstrated the internal test results of the o3 model. For example, in the field of software engineering, o3 achieved an accuracy rate of 71.7%, which is about 23 percentage points higher than the o1 model. Additionally, o3 scored 2,727 points on the competitive programming platform Codeforces. Furthermore, o3 outperformed the o1 model when answering mathematical and PhD-level scientific questions.
Notably, the o3 model reached a score of 87.5% on the ARC-AGI test, which compares AI performance with human intelligence. “A score of 85% for human performance is already of comparative value. Scores above this level are an important milestone,” said Greg Kamradt from the ARC Prize Foundation. This indicates that the new-generation model is closer to Artificial General Intelligence (AGI), capable of learning and acquiring new skills.
The o3 model scored 87.5% on the ARC-AGI test.
Image / OpenAI YouTube
In addition, o3-mini supports low, medium, and high computational modes, meaning users can adjust the model’s computation and thinking time based on the complexity of the task.
The new models have not yet been officially released. OpenAI CEO Sam Altman announced that o3-mini will be launched in January next year, followed by o3. Before January 10, o3-mini will be made available to cybersecurity researchers for testing the model’s safety.
### What Story Did OpenAI Tell in Its 12-Day Livestream at the End of 2024?
Before the official release of the o1 model, there had been industry speculation that the scaling law—using larger computational power, more parameters, and more data to improve model performance—was nearing its limits. OpenAI’s simultaneous release of the o1 and o3 models at the year’s end dispelled much of this pessimism. Additionally, the 12-day live broadcast featured three major highlights:
**Highlight 1: Developer-Friendly Customization Features**
In the past, large language models (LLMs) had mixed results when applied to highly specialized industries. OpenAI introduced a new feature, Reinforcement Fine-Tuning (RFT), which allows developers to create custom models tailored to professional domains such as law, healthcare, and finance with simple programming.
Moreover, OpenAI opened up the API for the o1 model, offering new functionalities such as function calling, structured outputs, developer messages, and visual inputs. Developer messages allow for specifying detailed instructions that the model should follow, such as defining tone and presentation style. The most intuitive application is that users can integrate the API to create personalized chat AI with their own style.
**Highlight 2: AI-Assisted Workflow**
This year, OpenAI introduced collaborative editing tools like Canvas and the AI search engine SearchGPT. Canvas enables users to invite AI to assist with editing, supplementing information, or making design suggestions during writing or programming tasks. SearchGPT integrates real-time web search into ChatGPT.
The new “Projects” feature combines Canvas and SearchGPT, allowing users to customize workspaces based on different tasks. Whether it’s basic file management, team collaboration, or integrating code and custom commands, all can be accomplished within the project space.
**Highlight 3: Multi-Modal AI Tools in Action**
In the series of live broadcasts, OpenAI once again showcased the trend of expanding language models from text to voice and images. Users can now interact with ChatGPT using voice, either through the online version or by making a phone call. ChatGPT Plus and ChatGPT Pro subscribers can also enable video functionality, allowing ChatGPT to interact with real-world surroundings.
OpenAI also officially launched the AI video generation model, Sora, which can create new video content from text, images, or existing videos. Sora supports up to 1080p resolution and generates videos of up to 20 seconds long. Users can also edit video content frame by frame.
On the first day of the broadcast, OpenAI introduced a $2,000 per month (approximately NT$65,443) ChatGPT Pro subscription plan. Subsequent new features, including Sora, were designed with differentiated plans for subscribers, further emphasizing OpenAI’s strategic shift toward a commercial model. Additionally, OpenAI has been actively integrating ChatGPT with Apple Intelligence and other applications, ensuring that by the end of the year, users could experience ChatGPT everywhere.
**Related Reading**: *OpenAI Launches o1 Model! Announces New ChatGPT Pro Plan with a Monthly Fee of Up to 6,500 New Taiwan Dollars—What’s So Special About It?*
**Editor-in-Chief**: Li Xiantai
