OpenAI最强编程智能体上线ChatGPT，AI的“终极外挂”来了

Core Insights - OpenAI has officially launched Codex, a cloud-based software engineering intelligent agent designed to assist developers in efficiently completing programming tasks [2][3] - Codex can handle various tasks including writing functional code, understanding code structure, answering questions about codebases, fixing bugs, and submitting reviewable pull requests [5][12] Group 1: Product Features - Codex features a simple user interface with an input box and two buttons: "Ask" and "Code," allowing users to describe tasks clearly for execution [3] - Each task is executed in an independent cloud sandbox environment, with runtime varying from 1 to 30 minutes based on complexity, and users can monitor progress in real-time [5] - The core model of Codex is codex-1, specifically fine-tuned for software engineering within OpenAI's o3 series, achieving a score of 72.1% on the SWE-Bench benchmark [8][9] Group 2: Training and Security - Codex employs reinforcement learning in real development environments, generating code that aligns with human coding styles and preferences while adhering strictly to instructions [12] - The system is designed with security in mind, only accessing provided codebases and preset dependencies, and it can identify and reject requests related to malicious development [16] Group 3: User Accessibility and Pricing - Codex is available to ChatGPT Pro, Enterprise, and Team users starting today, with support for Plus and Edu users expected soon [13][28] - For developers using the codex-mini-latest model, pricing is set at $1.5 per million input tokens and $6 per million output tokens, with up to 75% discount on prompt caching [29] Group 4: Market Position and Use Cases - OpenAI has been using Codex internally to enhance efficiency in tasks like refactoring and writing tests, receiving positive feedback from external partners like Cisco and Temporal [17] - Codex is particularly beneficial for experienced developers, enabling them to manage multiple tasks simultaneously, while it may not be as user-friendly for novice engineers [21][22] Group 5: Future Vision - OpenAI envisions Codex as more than just a coding tool; it aims to shape future collaboration models with features like multi-agent systems and asynchronous execution, potentially integrating into IDEs and Git tools [30][32]