OpenAI has been conducting a Christmas countdown live stream for 12 consecutive working days starting from December 4, U.S. time. On the first day, they launched the official version of model o1, and on the final day, they surprised everyone by announcing the latest generation models o3 and o3-mini.
### Highlights of ChatGPT’s Christmas Gift Package
Similar to o1, OpenAI’s latest models, o3 and o3-mini, also utilize the “Chain of Thought” (CoT) reasoning model technology, but with significantly improved performance compared to o1.
The performance of the o3 model has significantly improved over o1.
Image / OpenAI YouTube
During the live stream on December 20, U.S. time, OpenAI’s Vice President of Research, Mark Chen, showcased the internal testing results of the o3 model. For instance, in software engineering, o3 achieved an accuracy of 71.7%, which is about 23 percentage points higher than the o1 model; o3 also scored an impressive 2,727 points on the programming competition platform Codeforces. Additionally, o3 outperformed the o1 model in answering mathematics and PhD-level science questions.
Notably, the o3 model achieved a score of up to 87.5% in the ARC-AGI test, which aims to compare AI with human intelligence. “A score of 85% for human performance is of comparative value, and any score above that is a significant milestone,” said Greg Kamradt from the ARC Prize Foundation. This indicates that the new generation of models is closer to Artificial General Intelligence (AGI), capable of learning and acquiring new skills.
The o3 model can achieve a score of up to 87.5% in the ARC-AGI test.
Image / OpenAI YouTube
Moreover, the o3-mini supports low, medium, and high computation 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 officially launched, but OpenAI CEO Sam Altman stated that o3-mini is expected to be released in January next year, followed by o3. Before January 10, the o3-mini model will be available for application by cybersecurity researchers to assist in testing the model’s safety.
### What Story Did OpenAI Tell During the 12-Day Live Stream at the End of 2024?
Before the official release of the o1 model, there were several voices in the industry questioning whether the scaling law—trading larger computational power, parameters, and data volume for better model performance—was nearing its limits. OpenAI’s simultaneous announcement of both the o1 and o3 models at the end of the year seemed to dispel the pessimism. Additionally, the 12-day live stream featured three major highlights:
**Highlight 1: Developer-Friendly Customization Features**
In the past, large language models (LLMs) had mixed results when it came to leveraging highly specialized industry knowledge. OpenAI has introduced a Reinforcement Fine-Tuning (RFT) feature that allows developers to create customized models tailored for professional fields such as law, healthcare, and finance with simple programming. Furthermore, OpenAI has opened the API for the o1 model, providing features including function calling, structured outputs, developer messages, and visual input. Among these, developer messages can specify detailed instructions for the model to follow, such as defining tone and presentation style. The most intuitive application allows users to connect the API to create a chat AI with a personal style.
**Highlight 2: AI-Assisted Workflow**
This year, OpenAI has successively launched collaborative editing tools like Canvas and the AI search engine SearchGPT. Canvas allows users to invite AI to assist with editing, supplementing information, or providing design suggestions while writing or coding; SearchGPT integrates real-time web searches into ChatGPT. The newly launched “Projects” feature combines Canvas and SearchGPT, enabling users to customize their workspace based on different task requirements. Whether for basic file management, multi-person collaboration, or integrating code and custom commands, all can be achieved within the project space.
**Highlight 3: Multi-Modal AI Tools Come to Life**
Throughout the series of live streams, OpenAI showcased the trend of language models expanding from text to sound and images. Now, users can interact with ChatGPT via voice through the online version or by making a phone call. Subscribers of ChatGPT Plus and ChatGPT Pro can also enable video features, allowing ChatGPT to interact based on the real-world scene in front of them. OpenAI has also officially released the AI video generation model Sora, which can create new video content from text, images, or videos. Sora supports up to 1080p quality and generates videos lasting up to 20 seconds, with users able to edit image content frame by frame.
On the first day of the live stream, OpenAI launched a ChatGPT Pro subscription plan for $2,000 per month (approximately NT$65,443). Subsequent new features, including Sora, are designed with differentiated plans for subscribers, highlighting OpenAI’s strategic determination to shift towards a commercial model this year. Additionally, OpenAI is actively integrating ChatGPT with Apple Intelligence and other applications, allowing users to experience the omnipresence of ChatGPT by the end of the year.
### Further Reading: OpenAI Launches o1 Model! Releases Latest ChatGPT Pro Plan: Monthly Fee Up to NT$6,500, What’s So Impressive?
Editor: Li Xiantai