一文读懂GPT-5发布会：新功能乏善可陈

Core Insights - OpenAI has launched ChatGPT Agent, a universal AI agent that represents a significant advancement in AI technology [2][4][16] - The ChatGPT Agent integrates capabilities from previous projects, Operator and Deep Research, into a unified system that can perform complex tasks autonomously [16][20] - The agent operates in a virtual sandbox environment, allowing users to observe its actions in real-time while ensuring task context is maintained [18][23] Technical Features - ChatGPT Agent utilizes a new model trained through reinforcement learning, enabling it to switch seamlessly between various tools for complex tasks [20][21] - It includes a visual browser, text-based browser, terminal, and API access, allowing for diverse interactions and task execution [21][23] - The agent can autonomously execute tasks based on natural language commands, such as generating reports or analyzing data [23][24] Performance Metrics - In benchmark tests, ChatGPT Agent achieved a record score of 41.6% on the HLE task, surpassing the recently released Grok4 [26][27] - It demonstrated high accuracy in real-world data science tasks, with analysis and modeling accuracy rates of 89.9% and 85.5%, respectively [33] - The agent's ability to edit spreadsheets outperformed competitors, achieving a score of 45.5% in SpreadsheetBench [36] Availability and User Interaction - The ChatGPT Agent is available to Pro users immediately, with other tiers to follow in the coming days [42] - Users can interact with the agent by switching to "agent mode" and providing task descriptions, with real-time visibility into its operations [44] - The agent allows for periodic task execution, enhancing user convenience [45] Competitive Landscape - ChatGPT Agent's innovation lies in its integrated virtual machine environment, a feature not present in other models [51] - OpenAI acknowledges that while the agent is advanced, it is still in early stages, with ongoing improvements planned for functionality and user experience [53] - A comparison with Anthropic's Claude highlights differences in operational philosophy, with ChatGPT focusing on autonomous task execution within a virtual environment [54]