OpenAI officially announced on December 5 that it will hold a 12-day series of launch events, during which one product or sample will be released each day. The event will include the highly anticipated AI video generation tool Sora and a new reasoning model. This article will introduce the daily release content and related highlight information of the 12-day launch event.
Day 1: The full version of o1 launched, ChatGPT Pro subscription service
At the first day's launch event, OpenAI officially launched its most powerful reasoning modelo1the full version and its Pro version. The model is more complete than the previous o1-preview version, with performance improved by 34% and error rate reduced by 34%. It supports multimodal input. At the same time, OpenAI launched the ChatGPT Pro subscription service at $200 per month, providing unlimited access to o1 and the professional version of o1.
Highlight analysis:
- 01 Basic version: Suitable for daily use, powerful and responsive.
- Enhanced multimodal capabilities: The ability to process text and images is greatly improved, supporting image content analysis and answering related questions.
- Improved language understanding: More accurately captures language details and context, with smoother and more natural performance in text generation, translation, and summarization.
- Lower error rate: Compared with the preview version, the error rate is reduced by 35%, making generated content more accurate and reliable.
- Improved response speed: Speed is increased by 50%, reducing user waiting time and bringing a smoother experience.
- 01 Pro version: For professional users with higher requirements for AI performance, powerful and precise.
- Deep reasoning ability: Can handle complex mathematics and science problems, support multi-step reasoning and professional calculations, with thinking time of about 1-3 minutes.
- Professional domain knowledge: Performs excellently in fields such as science, mathematics, and programming, providing professional and accurate answers.
- Long text processing capability: Supports longer text input and output, easily handling long document processing and content generation.
- Customization capability: Provides more fine-tuning options, allowing users to personalize according to their own needs.
Original video:
Day 2: Reinforcement Fine-Tuning
At the second day's launch event, OpenAI introduced reinforcement fine-tuning technology. This is a new model training method that supports users in creating expert models in specific domains using a small amount of training data. Reinforcement fine-tuning optimizes the model's output through reinforcement learning algorithms, making it better match user expectations. Reinforcement fine-tuning is expected to be available to users only in spring 2025.
Highlight analysis:
- Training with a small amount of data: Using very few samples can enable the model to learn new reasoning methods.
- Reinforcement learning algorithm: Continuously optimizes the model's output based on the scoring mechanism provided by the user.
- Suitable for complex tasks: Especially suitable for fields requiring in-depth professional knowledge, such as law, finance, and healthcare.
Original video:
Day 3: The official version of Sora appears
At the third day's launch event, OpenAI launchedSorathe official version and the new version Sora Turbo. Sora is a video generation tool that can generate coherent and realistic video content based on text descriptions. Sora supports generating videos up to 1080p resolution and up to 20 seconds long, supports multiple video aspect ratios, and provides a variety of video editing functions. At this stage, Sora is only available toChatGPTPlusAvailable to Pro users, Plus users have a monthly quota of 50 video generations, while Pro users have up to 5000.
Highlight Analysis:
- Storyboard Feature: Users guide multiple actions in a video through a timeline, creating complex video sequences.
- Multiple Video Editing Features: Including Remix, Re-cut, Storyboard, etc., enhancing the logic and smoothness of creation.
- Video Mixing: Supports merging two video scenes into a brand-new scene, generating creative new content.
- High Resolution and Duration Support: Supports generating 1080p resolution and videos up to 20 seconds long.
Original Video:
Day 4: ChatGPT Canvas fully open, human+AI collaboration mode launched
At the fourth day's launch event, OpenAI announcedCanvasThe feature is fully open. Upgrading from a chat tool to a productivity tool, Canvas supports users collaborating with ChatGPT in writing and programming, providing a shared canvas where users and ChatGPT can jointly edit documents and code.
Highlight Analysis:
- Writing and Programming Collaboration: Users can complete writing and coding projects together with ChatGPT.
- Supports Python Code Execution: Canvas has almost all commonly used Python libraries built in, allowing real-time collaboration with AI while writing code.
- Image Recognition: Canvas supports image recognition; users can upload images and generate related text content.
- Canvas： Provides collaboration features for writing, reviewing, coding, and image recognition.
Original Video:
Day 5: OpenAI partners with Apple, ChatGPT integrated into Apple Intelligence
At the fifth day's launch event, OpenAI announced that ChatGPT is officially integrated into Apple Intelligence. This means iPhone, iPad, and Mac users can use ChatGPT's features through Siri.
Highlight Analysis:
- Siri Support: Users can complete more complex tasks through Siri voice commands.
- Writing Tool Upgrade: Apple's intelligent writing tools will introduce ChatGPT capabilities.
- Visual Intelligence Capabilities: Users can use the search function to quickly identify objects in front of them.
- ChatGPT Integration: Provides ChatGPT's intelligent support within the iOS, iPadOS, and MacOS ecosystems.
Original Video:
Day 6: OpenAI's "Her" launched, advanced voice model enhanced
At the sixth day's launch event, OpenAI brought video input and real-time screen sharing capabilities toChatGPT's advanced voice mode. This enables ChatGPT to "see" the user and the surrounding environment, and based on this visual information, better understand the user's instructions and questions. The advanced voice mode uses the multimodal 4o model, can directly input audio and output audio, supports more than 50 languages, is aimed at mobile users, and is available to Plus and Pro subscribers.
Highlight Analysis:
- Video Input Feature: ChatGPT's advanced voice mode has added a video input feature, enabling it to "see" the user.
- Screen Sharing Feature: Through the screen sharing feature, ChatGPT can help users reply to messages, assist with work and learning, and more.
- Santa Claus Virtual Character: Users can have video or voice calls with "Santa Claus" through the advanced voice mode.
Original Video:
Day7: ChatGPT launches the Projects feature
At the launch event on the seventh day, OpenAI introduced Projects new feature. It supports users in centralizing relevant materials, files, chat records, and other information into a single "Project," and supports scenarios such as project management and writing, file and data management, and personalization. It integrates Canvas and ChatGPT Search, is available to ChatGPTPlus/Pro users, and is planned to be offered to enterprise and education users early next year.
Highlight analysis:
- Project management: Users can create projects, upload files, and set personalized instructions.
- Canvas and ChatGPT Search integration: The Projects feature integrates Canvas and web search capabilities.
- Multi-step task execution: It lays the foundation for executing multi-step tasks in the future.
- Projects： Supports users in uploading files, setting custom instructions, grouping chats and data, and more.
Original video:
Day8: ChatGPT Search fully upgraded, free for everyone
At the launch event on the eighth day, OpenAI showcasedChatGPT Searchthe latest search feature, based on a fine-tuned version of the GPT-4o model. By combining content provided directly by third-party search providers and ChatGPT's partners, the feature enables users to quickly and accurately obtain the information they need. The newly searchable information includes real-time information such as the latest sports scores, news, stock quotes, and more. Users can get fast, timely answers through links pointing to relevant web resources.
Highlight analysis:
- Improved search performance: Search is faster, the mobile experience is better, and a map experience has been added.
- Voice search integration: Users can obtain real-time web information through voice conversations.
- Free worldwide access: The search feature is free for all logged-in users worldwide.
- Conversational search: ChatGPT can maintain conversation context and understand user intent.
Original video:
Day9: o1 reasoning model opens API, costs plunge 60%
At the launch event on the ninth day, OpenAI announced the official API release of the o1 model, with costs reduced by 60%, customizable upgrades, and new features such as function calling, developer messages, and image analysis. It is an important technological upgrade. The new model is not only smarter, but can also respond to user needs faster. The Realtime API has been fully upgraded, including the introduction of WebRTC support, lower prices, and more flexible response control.
Highlight analysis:
- o1 model API: Officially launched the full-power o1 model API, supporting features such as function calling, developer messages, Structured Outputs, and visual recognition.
- Realtime API upgrade: Integrated simple WebRTC functionality, and GPT-4o audio pricing was reduced by 60%.
- Preference fine-tuning: Introduced preference fine-tuning based on the Direct Preference Optimization (DPO) algorithm.
- Multilingual SDK support: Opened beta versions of the Go and Java SDKs, supporting multiple programming languages.
- Performance improvement: The o1 model has demonstrated excellent performance in multiple fields, such as achieving a high score of 76.6% in programming ability tests.
Original video:
Day10: Use ChatGPT via phone and WhatsApp
At the Day 10 launch event, OpenAI introduced new features for using ChatGPT via phone and WhatsApp. This update means users can converse with the most advanced AI through the most basic phone functionality or WhatsApp, at the number 1-800-CHAT-GPT, or 1-800-242-8478. It supports devices such as smartphones, feature phones, and landlines. This greatly lowers the barrier to use.
Highlight analysis:
- Phone access: Users can have voice conversations with ChatGPT by dialing 1-800-CHAT-GPT.
- WhatsApp integration: Users worldwide can message ChatGPT on WhatsApp, and the model providing the service is 4o-mini.
- Simple and easy to use: No need to register an account, no need to remember a password, just operate as simply as making a phone call.
- Cross-device support: From the latest iPhone to old flip phones, all can directly converse with the AI.
- Multilingual support: ChatGPT can provide multilingual services, such as teaching users to communicate in Spanish.
Original video:
Day11: Deep integration of ChatGPT with Mac apps
In the Day 11 livestream, OpenAI announced deep integration of ChatGPT with Mac apps, with updates especially in programming and writing. Now, users can work with apps such as Warp and XCode through simple copy-and-paste operations, executing tasks without detailed communication. In addition, ChatGPT also supports working with apps such as Notion and Apple Notes in voice mode. These new features have been rolled out in the latest version of the ChatGPT Mac app, while Windows users still need to wait.
Highlight analysis:
- Brand-new upgrade to the desktop app: The ChatGPT desktop app supports working with a variety of applications, including development tools and document editing tools. ChatGPT analyzes Git commit data and generates visual charts.
- Real-time demonstration of intelligent collaboration with the Warp terminal and Xcode IDE: It demonstrated how ChatGPT seamlessly collaborates with the development environment, providing real-time support for programming tasks. It achieves code autocompletion and feature addition in Xcode.
- Support for writing tools such as Notion, Apple Notes, and Quip: These features are now available in the Mac desktop app, and Windows users need to wait.
- Advanced voice interaction mode: Talking with ChatGPT through voice mode provides a more natural and efficient experience.
- Search assistance feature: Ensures the accuracy of output information, combining search functionality to optimize writing content.
Original video:
Day12: Introduction to the release of the latest-generation reasoning models o3 and o3-mini
OpenAI released its strongest reasoning modelo3, claiming it is close to artificial general intelligence (AGI). o3 scored 87.5% on the ARC-AGI test, far exceeding GPT-3 and GPT-4o. o3 has a full version and a mini version, and the mini version will be launched at the end of January. o3 scored 2727 in the programming competition Codeforces, and achieved 96.7% accuracy on the math benchmark AIME 2024. OpenAI will collaborate with the ARC-AGI Foundation to build the next benchmark test.
Highlight analysis:
- Upgraded mathematical reasoning ability: o3 performs excellently on complex math problems, reaching 96.7% accuracy in the American AIME math competition, demonstrating the problem-solving ability of a top mathematician.
- Outstanding programming performance: It obtained an ELO score of 2727 on the CodeForces programming competition platform, surpassing the level of top programmers.
- Scientific problem-solving ability: On the GPQA science benchmark, o3 achieves 87.7% accuracy, surpassing the average human expert level (70%).
- Transparent reasoning path: Provides a clear reasoning process, able to show the logic and intermediate conclusions of each step.
Original video: