AI News AI资讯 8h ago Updated 6h ago 更新于 6小时前 48

Understanding ChatGPT Work 理解 ChatGPT Work

ChatGPT Work is split into two products: Work Cloud (accessible via chatgpt.com and mobile apps) and Work Local (desktop app with file/program access), with the article focusing on Work Cloud Work Cloud offers unique capabilities absent in regular Chat: code execution with full internet access, a headless Chrome browser with JavaScript execution, persistent filesystem across sessions, sub-agent delegation, and model selection (GPT-5.6 Sol/Luna/Terra with reasoning levels) The code execution envi OpenAI于2026年7月9日发布ChatGPT Work,分为Work Cloud(云端)和Work Local(本地桌面应用)两个版本,仅对$20/月及以上订阅用户开放 Work Cloud核心差异化功能包括:支持GPT-5.6 Sol/Luna/Terra多模型选择、带互联网访问的代码执行环境、无头Chrome浏览器工具、跨会话持久文件系统 代码执行环境突破性地支持全互联网访问(默认开放),可克隆GitHub仓库、安装依赖并与外部API交互,远超Claude受限的白名单模式 Work与Chat的计费体系分离(Work使用Codex配额,Chat使用独立配额),解释了两者模型选择差异的原

72
Hot 热度
70
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • ChatGPT Work is split into two products: Work Cloud (accessible via chatgpt.com and mobile apps) and Work Local (desktop app with file/program access), with the article focusing on Work Cloud
  • Work Cloud offers unique capabilities absent in regular Chat: code execution with full internet access, a headless Chrome browser with JavaScript execution, persistent filesystem across sessions, sub-agent delegation, and model selection (GPT-5.6 Sol/Luna/Terra with reasoning levels)
  • The code execution environment's internet access is a major differentiator, allowing cloning of GitHub repos, package installation, and unrestricted web interaction—far surpassing Claude's restricted allowlist approach
  • Work is gated behind $20/month subscriptions and billed against a separate Codex allowance, distinct from Chat's allowance, which explains model availability differences
  • The persistent /workspace/scratch filesystem is shared across all running Work sessions, enabling cross-session file access but not shared process space or localhost server access

Why It Matters

ChatGPT Work represents OpenAI's strategic push to position ChatGPT as a task-completion engine rather than just a conversational interface, directly competing with Claude's coding and browser automation capabilities. The introduction of unrestricted internet-accessible code execution in a cloud environment is a significant capability leap that closes a gap many AI practitioners had noted since the original Code Interpreter launch. Understanding the Work vs. Chat distinction and the underlying billing/model architecture is essential for professionals evaluating which OpenAI product tier delivers the most value for their workflows.

Technical Details

  • Architecture split: Work Cloud runs entirely in OpenAI's infrastructure (accessible via web/mobile), while Work Local runs on the user's machine via the desktop app (formerly Codex), with file and program access
  • Model selection: Work offers GPT-5.6 Sol, Luna, and Terra with reasoning levels (Light/Medium/High/Extra High/Max/Ultra), plus GPT-5.5 at Light/Medium/High/Extra High; Chat offers a separate lineup including 5.6 Instant and 5.6 Pro (exclusive to Chat, $100+/month tier)
  • Code execution with internet: The sandboxed code environment supports full internet access by default (configurable domain allowlists), enabling GitHub repo cloning, PyPI/NPM package installation, and arbitrary web API interaction—unlike Claude's short allowlist
  • Headless Chrome browser: Built on Playwright, supports form filling, screenshots, JavaScript execution against the DOM, and secure credential handling (user enters passwords/2FA directly without model exposure)
  • Persistent filesystem: Each Work session gets a /workspace/scratch/ folder that persists across sessions and is mounted across all running Work sessions simultaneously, though process space and localhost servers remain isolated per session
  • Sub-agent delegation: Ultra reasoning mode eagerly delegates to sub-agents; Work supports running sub-agent sessions with Sol, Luna, and Terra models
  • Access control: Available only to $20/month subscribers and above; free and $8/month Go users are excluded

Industry Insight

  • OpenAI is clearly segmenting its product stack along a "chat for information, work for execution" paradigm—practitioners should evaluate whether their workflows justify the $20/month upgrade based on the unique capabilities (internet-connected code execution, persistent storage, browser automation) rather than treating Work as merely a rebranded Codex
  • The unrestricted internet access in the code execution sandbox represents a significant capability advantage over competitors like Claude, but also raises security and abuse considerations that may prompt OpenAI to introduce tighter default restrictions in future iterations
  • The separate billing allowance for Work vs. Chat suggests OpenAI is experimenting with differentiated cost structures for high-compute tasks (sub-agent delegation, browser sessions, code execution)—professionals should monitor how these allowances evolve as the product matures and consider the total cost of ownership for heavy Work usage

TL;DR

  • OpenAI于2026年7月9日发布ChatGPT Work,分为Work Cloud(云端)和Work Local(本地桌面应用)两个版本,仅对$20/月及以上订阅用户开放
  • Work Cloud核心差异化功能包括:支持GPT-5.6 Sol/Luna/Terra多模型选择、带互联网访问的代码执行环境、无头Chrome浏览器工具、跨会话持久文件系统
  • 代码执行环境突破性地支持全互联网访问(默认开放),可克隆GitHub仓库、安装依赖并与外部API交互,远超Claude受限的白名单模式
  • Work与Chat的计费体系分离(Work使用Codex配额,Chat使用独立配额),解释了两者模型选择差异的原因

为什么值得看

本文深入剖析了OpenAI最新企业级产品ChatGPT Work的技术架构与功能边界,为AI从业者理解OpenAI在Agent化、工具链整合方面的战略方向提供了第一手实践洞察。其代码执行环境开放互联网访问的设计,标志着AI助手从"对话工具"向"任务执行平台"的关键跃迁。

技术解析

  • 双版本架构:Work Cloud通过chatgpt.com和移动端访问,Work Local为桌面应用(原Codex应用重命名),后者可访问本地文件和运行程序,定位更偏向非开发者用户
  • 模型选择体系:Work支持GPT-5.6 Sol/Luna/Terra(含Light至Ultra共7档推理级别)及GPT-5.5,与OpenAI API模型一致;Chat则提供5.6 Instant/Pro等独立选择,$20用户最高仅到High级别
  • 代码执行环境:突破性地开放互联网访问权限,可自由克隆GitHub仓库、安装PyPI/NPM包、调用外部API,默认允许全域名访问(可配置白名单),相比Claude的短白名单模式大幅扩展了能力边界
  • 无头Chrome浏览器:支持完整Chrome实例启动、网页加载、表单填写、截图,可执行JavaScript操作DOM,登录场景支持用户接管输入密码和2FA而不经过模型,实现安全凭证隔离
  • 持久文件系统:每个Work会话拥有独立scratch文件夹(如/workspace/scratch/xxx),跨会话持久化且可被多个运行中的Work会话共享挂载,实现文件级协作但进程空间隔离

行业启示

  • AI产品分层策略深化:OpenAI通过Work/Chat分离、Codex配额独立计费,构建了从免费用户到$100/月Pro用户的精细化变现矩阵,提示行业需关注AI产品的订阅层级设计与功能 gating 策略
  • Agent化基础设施竞争白热化:代码执行+互联网访问+浏览器自动化+持久存储的组合,标志着AI助手正从"问答引擎"演变为"可执行任务的智能体平台",Claude、Gemini等竞品需加速补齐类似工具链能力
  • 安全与能力的平衡点探索:Work允许用户接管登录流程而非让模型处理凭证,体现了对敏感操作的安全设计考量;但代码执行环境默认开放全互联网访问仍需谨慎评估,行业需建立相应的使用规范与审计机制

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPT GPT OpenAI OpenAI Conversational AI 对话系统 LLM 大模型 Product Launch 产品发布