Operator 由计算机使用代理（CUA）驱动，这是一种结合 GPT-4o 视觉能力与强化学习高级推理的模型。CUA 经过训练，能够像人类一样与图形用户界面（GUI）交互，从而无需操作系统或网络专用 API 即可灵活执行数字任务。它建立在多模态理解与推理交叉领域多年基础研究之上，通过结合高级 GUI 感知与结构化问题解决，能够将任务分解为多步计划，并在遇到挑战时自适应地自我纠正。这一能力标志着人工智能发展的下一步，使模型能够使用人类日常依赖的相同工具，为广泛的新应用打开大门。
Powering Operator with Computer-Using Agent, a universal interface for AI to interact with the digital world. Today we introduced a research preview of Operator(opens in a new window), an agent that can go to the web to perform tasks for you. Powering Operator is Computer-Using Agent (CUA), a model that combines GPT‑4o's vision capabilities with advanced reasoning through reinforcement learning. CUA is trained to interact with graphical user interfaces (GUIs)—the buttons, menus, and text fields people see on a screen—just as humans do. This gives it the flexibility to perform digital tasks without using OS-or web-specific APIs. CUA builds off of years of foundational research at the intersection of multimodal understanding and reasoning. By combining advanced GUI perception with structured problem-solving, it can break tasks into multi-step plans and adaptively self-correct when challenges arise. This capability marks the next step in AI development, allowing models to use the same tools humans rely on daily and opening the door to a vast range of new applications.
核心贡献 · Key contributions
提出计算机使用智能体（CUA），结合 GPT-4o 视觉能力与基于强化学习的推理，实现图形用户界面交互。 Introduces Computer-Using Agent (CUA) combining GPT-4o vision with RL-based reasoning for GUI interaction.
使用统一的屏幕-鼠标-键盘接口，在 OSWorld（38.1%）、WebArena（58.1%）和 WebVoyager（87%）上取得最先进结果。 Achieves state-of-the-art on OSWorld (38.1%), WebArena (58.1%), and WebVoyager (87%) using a universal screen-mouse-keyboard interface.
展示了测试时缩放：允许更多步骤时性能提升。 Demonstrates test-time scaling: performance improves with more allowed steps.
实施分层安全缓解措施，应对滥用、模型错误和前沿风险。 Implements layered safety mitigations against misuse, model mistakes, and frontier risks.
CUA 能自适应自我纠正并处理多步任务，无需操作系统或网络专用 API。 CUA can adaptively self-correct and handle multi-step tasks without OS- or web-specific APIs.
以 Operator 研究预览形式向美国 Pro 用户发布，收集真实世界反馈。 Released as Operator research preview for Pro users in the U.S., enabling real-world feedback collection.
局限 · Limitations
在 WebArena 等复杂基准上性能（58.1%）仍落后于人类水平（OSWorld 上 72.4%）。 Performance on complex benchmarks like WebArena (58.1%) still lags behind human level (72.4% on OSWorld).
CUA 可能犯错，如打字错误、错误购买或意外删除，需要用户确认。 CUA may make mistakes such as typos, wrong purchases, or accidental deletions, requiring user confirmations.
易受提示注入和越狱等对抗性攻击，尽管已有缓解措施。 Vulnerable to adversarial attacks like prompt injections and jailbreaks, though mitigations are in place.
目前仅限于美国 Pro 用户，并阻止某些高风险任务（如银行交易）。 Currently limited to U.S. Pro users and blocks certain high-risk tasks (e.g., banking transactions).
早期技术，并非在所有场景下可靠，尤其是敏感决策任务。 Early-stage technology; not reliable in all scenarios, especially sensitive decision-making tasks.