计算版图正转向具备直接屏幕交互与任务执行能力的智能体模型。GPT-6 Astra、Claude Fable 5.1等新工具通过导航软件界面来自动化复杂工作流。
核心要点：OpenAI的GPT-6 Astra被定位为业界领先的computer-use模型，专注浏览、软件工程与专业自动化；Anthropic的Claude Fable 5.1面向知识工作与长文档分析，提供高级智能体能力；Meta的Muse通过“Muse Secure VM”运行，用一个专用虚拟机同时承载Agent与用户数据以保障隐私；HeyClicky则是macOS专属助手，通过用户触发热键获取屏幕上下文。
桌面智能体AI的兴起：从聊天走向“用电脑”
个人计算的版图正进入一个由“能直接看屏幕、能执行任务”的智能体模型定义的阶段。OpenAI的GPT-6 Astra与Anthropic的Claude Fable 5.1等头部实验室的最新发布，标志着行业越过聊天界面，迈向可以自主导航软件、填写表单、管理复杂工作流的桌面智能体。
这代表了人机交互的根本性转变：模型不再要求用户通过API或手工输入逐个操作应用，而是使用computer-use能力解读界面上的视觉元素，并代表用户执行动作。这一转变的目标，是通过自动化琐碎的行政与技术工作流，来解决高延迟的专业任务。
技术实现与模型规格
这一代桌面智能体采用了不同的架构路径来获得屏幕感知与任务自主性。GPT-6 Astra（OpenAI）在设计上是业界领先的computer-use模型，覆盖浏览、软件工程与专业自动化，并对齐性极高——按OpenAI系统卡片的说法，在安全评估中“超出预期范围”的比率为0%。
Claude Fable 5.1（Anthropic）针对知识工作与长文档分析优化，具备高级智能体能力，并包含一个专门针对网络安全与生命科学增强安全防护的Mythos变体。
Muse（Meta）通过“Muse Secure VM”运行——一个把Agent与用户数据一同安置的专用虚拟机，使用Muse Spark模型完成预订预约、处理客服等真实世界任务。HeyClicky是macOS专属助手，通过用户触发的热键读取屏幕上下文，可在屏幕上绘制视觉引导，并可按语音指令派出Agent执行任务。
性能基准
这些模型的效能，越来越多地以其在数字环境中导航的能力来衡量，而非只看文本生成。在computer use与推理维度，GPT-6 Astra在FrontierMath Tier 4取得98%、在ARC-AGI-3取得99.9%（据OpenAI）。在编码与终端任务维度，Terminal-Bench 4.0上GPT-6 Astra达到57.9%的准确率，超过Claude Fable 5.1的55.8%（据OpenAI研究）。在效率维度，OSWorld 2.0基准上，Astra完成任务比GPT-5.6 Sol的前代迭代快约47%（据OpenAI）。
可用性
GPT-6 Astra可通过ChatGPT Plus、Pro、Business与Enterprise各档使用，也开放给OpenAI API、Microsoft Azure与AWS Bedrock（据OpenAI博客）。Claude Fable 5.1面向付费用户及Anthropic API开放（据Anthropic）。Meta的Muse正在iOS、Android与muse.ai上逐步推出（据Meta）。
本文来源：https://www.thenextgentechinsider.com/pulse/agentic-desktop-ai-marks-a-shift-toward-autonomous-computer-use-models
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
The landscape of computing is shifting toward agentic models capable of direct screen interaction and task execution. New tools like GPT-6 Astra and Claude Fable 5.1 are automating complex workflows by navigating software interfaces.
Key Points: OpenAI's GPT-6 Astra is positioned as the leading computer-use model for browsing, software engineering, and professional automation; Anthropic's Claude Fable 5.1 focuses on knowledge work and long-document analysis with advanced agentic capabilities; Meta's Muse operates via a secure VM to handle both agent and user data privately; HeyClicky is a macOS-specific assistant triggered by custom hotkeys.
The Rise of Desktop Agent AI: From Chat to "Using the Computer"
Personal computing is entering a new phase defined by agentic models that can "see" the screen and execute tasks autonomously. The latest releases from top labs — including OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 — mark a departure from passive chat interfaces toward active desktop agents capable of navigating software, filling forms, and managing complex workflows.
This represents a fundamental shift in human-computer interaction: instead of requiring users to manually feed commands to apps, these models use computer-use capabilities to interpret visual elements on the interface and perform actions on the user's behalf. The goal is to automate tedious administrative and technical tasks — resolving high-latency professional workflows that currently slow down decision-makers.
Technical Architectures and Model Specifications
The latest generation of desktop agents employs diverse architectures to achieve screen awareness and autonomy. GPT-6 Astra (OpenAI) is designed as the industry's leading computer-use model, covering browsing, software engineering, and professional automation — with strong alignment, as evidenced by its 0% "outside-expected-range" failures in safety assessments (per OpenAI's system card).
Claude Fable 5.1 (Anthropic) is optimized for knowledge work and long-document analysis, offering advanced agentic capabilities alongside a "Mythos" variant tailored for cybersecurity and life sciences with enhanced safety.
Muse (Meta) operates via a "Muse Secure VM" — a dedicated virtual machine hosting both the agent and user data to ensure privacy — powered by the Muse Spark model for real-world tasks like booking appointments and handling customer service. HeyClicky is a macOS-specific assistant that activates via user-triggered hotkeys to read screen context, drawing visual cues on-screen and deploying agents per voice instruction.
Performance Benchmarks
The efficacy of these models is increasingly measured by their ability to navigate digital environments, not just generate text. In computer-use and reasoning, GPT-6 Astra achieves 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3 (per OpenAI). For coding and terminal tasks, it outperforms prior iterations on Terminal-Bench 4.0 with 57.9% accuracy — surpassing Claude Fable 5.1's 55.8% (per OpenAI research). In efficiency, OSWorld 2.0 benchmarks show Astra completing tasks ~47% faster than GPT-5.6 Sol (per OpenAI).
Availability
GPT-6 Astra is available through ChatGPT tiers (Plus / Pro / Business / Enterprise) and via API on OpenAI, Microsoft Azure, and AWS Bedrock (per OpenAI blog). Claude Fable 5.1 is accessible to paid subscribers and via Anthropic API (per Anthropic). Meta's Muse is rolling out on iOS, Android, and muse.ai (per Meta).