ChatGPT Work

• 1. overview (553 chars)

  OpenAI officially launched ChatGPT Work on July 9, 2026 - an autonomous working agent built into ChatGPT, driven by the flagship model GPT‑5.6 released on the same day. It transforms ChatGPT from a question and answer tool to an "execution-oriented" work platform that can autonomously execute long-term multi-step tasks across applications, files, and desktops. It also launches Sites site functions, scheduled tasks, and redo desktop applications. It was seen as a direct challenge to Anthropic Claude Cowork on the first day of its release, and also marked OpenAI's strategic transformation from a conversational assistant to a work execution tool.

• 2. Product background (787 chars)

  ChatGPT Work is the first step in OpenAI’s vision of “going beyond Q&A and turning big ideas into reality.” The bottom layer is driven by GPT‑5.6 and integrates the capabilities of the original Codex programming agent - more than 5 million people use Codex every week. Although it was originally a programming agent built for developers, more than 1 million people now use it for work scenarios other than software development, indicating that this set of capabilities can fully support more general tasks. GPT‑5.6, which was released at the same time, was fully launched after about two weeks of security review by the U.S. government. It adopts a three-level hierarchical system of "celestial body naming" (Sol sun / Terra earth / Luna moon). Token efficiency is increased by about 54% compared with the previous generation GPT‑5.5. The launch of ChatGPT Work is OpenAI’s positive response to Anthropic in the field of office automation.

• 3. Product functions and usage experience (2157 chars)

  The core of ChatGPT Work is to break down complex goals into small steps and complete them independently. Even if it takes several hours, it can continue to follow up. If there is any uncertainty, it will ask the user. The user can also adjust the direction and approve key actions at any time. To get started, you need to connect the tools where you work daily through plug-ins: Slack, Microsoft Teams, Google Drive, SharePoint, email, calendar, CRM, project progress tracker, etc. The new unified plug-in directory brings plug-ins together in one place. You can enter "@appname" in the prompt to directly specify the data source. ChatGPT will also recommend related plug-ins during the conversation.   In terms of output, it captures information from connected data sources, generates documents, presentations, spreadsheets and analytical reports, and continuously polishes drafts in the background. The Sites function (public beta) launched simultaneously this time can convert work results into interactive web applications and share them with the team or the public through URLs - real-time data panels, project progress tracking, release calendars, prototypes, internal portals and interactive reports can be generated directly and automatically updated when the underlying information changes. Scheduled Tasks support one-time, periodic repetition, event triggering and continuous monitoring, such as updating the site at 8 a.m. every day, refreshing the meeting agenda to review new Slack updates every week, and automatically updating presentations upon receiving emails containing new feedback.   The desktop side is the highlight of this update: the new ChatGPT desktop application has a built-in browser, which brings together websites, tools and online files in one place; Computer Use can simulate manual clicking, typing and moving files, and perform tasks across applications, tools and browsers in the background. The original independent Codex application was merged into the desktop, bringing inline diff editing, sidebar PR review, multi-repository support and faster Computer Use; the old desktop was renamed ChatGPT Classic; the Chrome extension was also updated to be available in the sidebar. Compared to Claude Cowork, which Anthropic added writing capabilities at the same time, ChatGPT Work is broader in multi-agent parallelism (concurrent subtasks) and Sites-unique features, but its deep integration with Microsoft 365 (OneDrive/SharePoint/Outlook reading and writing) is still not as mature as Cowork.

• 4. Pricing and business model (1124 chars)

  In terms of the pace of package opening, the web and mobile terminals will be open to Pro, Enterprise, and Edu users in the first batch, and will be expanded to Plus and Business in the next few days; the three major functions of Chat, Work, and Codex on the desktop will be open to all packages (including free users). The subscription tiers in July 2026 are: Free (limited), Go ($8/month, basic Work access), Plus ($20/month, full Work access), Pro ($100/month 5× or $200/month 20×, Ultra mode), Business ($20/seat/month, annual payment), Enterprise (customized). Work and Codex are both included in the paid package and do not need to be purchased separately.   The measurement method is worth noting: ChatGPT Work consumes the package quota according to the complexity of the task, and heavier tasks may fill up the subscription quota faster; Enterprise and Edu administrators can set expenditure control in the management console to achieve precise control through workspace-level default values, group restrictions, personal exception rules and review of top-up applications. GPT‑5.6’s API three-tier pricing is Sol $5/$30, Terra $2.5/$15, and Luna $1/$6 (input/output per million tokens). The flagship range is cheaper than Claude Opus 4.8, and the mid-range Luna is also more competitive than the entry price of Claude Sonnet 5.

• 5. User reviews (1540 chars)

  The positive feedback was overwhelming. Nearly 100% of OpenAI’s internal teams (including finance and sales) are already using ChatGPT Work and Codex to improve efficiency: the sales process turns an exploratory conversation into a customized proof of concept (PoC) in less than 24 hours (originally taking weeks); financial month-end closing and forecasting are reduced from days to hours. In external customer cases, Zapier uses Work to build a repeatable system to review thousands of potential customers every month, generate weekly executive data panels, and reveal seven-digit potential sales; Virgin Atlantic uses it to automate GTC preparations, replacing the Excel process that accounts for 40% of pre-meeting time; RingCentral uses it to compare differences in passenger experience in the airline industry, reducing weeks of analysis to hours; NVIDIA uses it to track customer registration, plan meetings, and integrate hundreds of meeting minutes for review after the meeting. Huxiu’s actual measurement believes that the delivery completion of GPT‑5.6 Sol plus Work is generally higher than that of Claude Cowork. With multi-modal capabilities such as Sites and graph generation, the $20 Plus gets more than the price.   The shortcomings are equally obvious. The most prominent one is quota anxiety: the agent quota experience is prone to overturning, and the subscription quota is used up in just a few heavy tasks. The few dollars of temporary recharge are also bottomed out in an instant, and the bill is hidden outside the subscription and is not transparent enough; Huxiu also pointed out that it is "slower and burns more tokens." Since it has only been online for 1 day, it can only evaluate early signals and does not yet have a stable reputation. Even the comparison PPT it generates truthfully states this.

• 6. Industry evaluation (774 chars)

  Many media regard ChatGPT Work as a key step in OpenAI's transformation from a conversational assistant to an execution work tool, directly benchmarking Anthropic. Sina Finance called the transition from Coding agents to Working agents a “direct challenge to Anthropic” and mentioned that the full launch of GPT‑5.6 underwent a two-week strict security review by the U.S. government. Third-party comparisons generally believe that ChatGPT Work has a lower entry barrier - Go $8 for basic Work, while Claude Cowork requires Max ($100+/month); however, chatgptaihub pointed out that Cowork has about six months of first-mover advantage and a more mature partner ecosystem, while Work counters by bridging one-end applications and SaaS. The overall judgment is that the two are more complementary than competitive: Work is used for Web and output tasks, and Cowork is used for local files and Microsoft 365 processes.

• 7. Controversies and Risks (921 chars)

  The first is that quotas and billing are not transparent: complex tasks are measured by volume, and bills are not included in subscriptions. Enterprises need to rely on management consoles for precise control, otherwise they are prone to overspending. The second is security governance: ChatGPT Work is built on Enterprise's security, privacy, compliance and workspace management. Auto-review will use advanced models to review key operations involving connected tools and APIs to prevent the leakage of sensitive information; the desktop side introduces enterprise-level protection (including agent network access policies). The third is the inherent risk of autonomous agents operating files - referring to the two incidents of Claude Cowork accidentally deleting 11GB of user files in the first week of its launch and being prompted for injection attacks by demonstrations, Work's Computer Use cross-application operations also require approval guardrails. In addition, the independent Atlas browser will gradually be offline, and users need to smoothly migrate to ChatGPT.

• 8. Applicable groups and usage suggestions (719 chars)

  Suitable for: Sales, marketing, finance, operations, and other teams that need to integrate information across applications and produce documents, reports, and demonstrations; teams that want to use Sites to quickly build internal dashboards, tracking tools, or prototypes; knowledge workers who already use Plus and above and want to take over the workflow with one command.   Not suitable or need to be cautious: For those who rely heavily on Microsoft 365 writing operations (Outlook/OneDrive/SharePoint) and require deep integration, Claude Cowork is currently more mature; teams that are sensitive to quota costs and want to run long-term tasks on a large scale but are unwilling to finely control bills must first control their expenditures. Alternatives include Anthropic Claude Cowork, Microsoft 365 Copilot, and Google Gemini Workflow.

• 9. Summarize (407 chars)

  ChatGPT Work is a key step for OpenAI to transform ChatGPT from "talking" to "doing": with GPT-5.6 as the engine, Sites and scheduled tasks as the highlights, and desktop Computer Use to connect the local, it initially has the independent work ability to benchmark Claude Cowork. Its shortcomings lie in the depth and credit billing transparency of Microsoft 365. It is suitable as the default agent for web and output tasks, and can be used with Cowork for heavy file and local processes.

No in-depth report yet

• 

  bigfrog122

  —

  Scheduled Tasks 这个功能挺实用的，设置每周五下午自动把 Slack 这一周的消息汇总成周报，到点直接收邮件就行，不用再手动整理了，省了不少重复劳动的时间。

---

• 

  LRodriguez_99

  —

  桌面端三视图切换目前还有点割裂感，Chat 里的对话不会自动同步到 Work，本地桌面文件也留在本机，换台电脑就得重来，这个体验还有待打磨，希望后续版本能改进。

---

• 

  HaroldStephensIII

  —

  Sites 功能生成网页确实快，描述了一下团队项目进度看板的需求，几分钟就出了个能用的页面，预览满意后一键发布就有公网链接了，小团队内部用够方便，连运维都不用麻烦。

---

• 

  crazypanda502

  —

  定时任务救我狗命。

---

• 

  AndreaPerez_728

  —

  贵，但好用。

---

• 

  DAOthinker340

  —

  回不去了，Work 真香。

---

• 

  ATorres_2022

  —

  额度掉得比我工资还快，Pro 20倍也顶不住。

---

• 

  Joe_KellyJr319

  —

  Reddit 上吵翻了，有人说这是 breakthrough 有人说是一团乱。说突破的是觉得 agent 终于能主动干活不用人盯着了，说乱的是觉得产品结构太分裂，三种模式加 Classic 把人搞晕了，说实话两种说法都有道理。

---

• 

  ryyvlya3_x

  —

  早上让它抓过去 24 小时行业要闻出简报，喝咖啡的功夫就有了，比我自己刷强。

---

• 

  beautifulgoose733

  —

  有开发者说给了 Sol Ultra 两个真正复杂的任务，结果一个任务就把使用额度全部吃完了，这定价模型确实让人心里没底，企业批量部署之前不搞清楚用量根本不敢下手采购。

---

• 

  AdamEdwards_Pro855

  —

  新 App 默认进 Work 模式，Chat 被塞进侧边栏子菜单，我也遇到一样的问题。从原生 Mac 改成 Electron 套壳之后，风扇转得比以前凶，开多个长任务明显卡。不过模型选择器做成 Faster 到 Smarter 的滑块确实省心，不用去记 Sol、Terra、Luna 这些天体名了。

---

• 

  M_etaHub

  —

  说实话 ChatGPT Work 的交付质量比我预期高。我让它把客户调研转成营销简报再改给不同市场，出来的东西结构完整、背景也连贯。但槽点也明显：跑几个重任务就把我 Pro 20 倍账号的额度烧穿了，临时充的几美元瞬间见底，账单还藏在订阅外不够透明。OpenAI 真该把用量说明写清楚。

---

• 

  MakaylaLopez

  —

  用 Work 把每周 Slack 动态整理成会议议程，现在周一早上一打开就有现成简报，省了我至少两小时。就是跑一次定时任务额度肉眼可见地往下掉。

---

• 

  KChavezIII98

  —

  关于计费不透明，OpenAI 员工说不同推理档 token 消耗差 5 到 10 倍，建议按难度选档。问题是默认设置会把人「不经意」推向更贵的 Sol，好多人发现用量比预期烧得快得多。他们罕见地两次重置用户限制、承诺恢复侧边栏，但根上的额度机制还得改。

---

• 

  EEvans_2020

  —

  Work 和 Codex 到底啥区别？官方说只是 UI 皮肤不同，懵。

---

• 

  樱花189

  —

  我们财务用 Work 做月末关账，原来要几天现在几小时，报表自动拉 Excel 核对。组内都惊了，说这钱花得值。

---

• 

  Gary_Brooks_2020

  —

  Sites 是真方便，一条指令出个内部门户，URL 一发就能看。

---

• 

  THart_X

  —

  看到 Matt Shumer 说 GPT-5.6 Sol 最高推理档把 Mac 上大部分文件删了，后背发凉。Work 的 Computer Use 能模拟点击、移动文件，这种跨应用操作真得把审批护栏拉满。Claude Cowork 上线首周也误删过 11GB，自主智能体操作本地文件就是高危，官方得把自动审核做扎实。

---

• 

  ALwag

  —

  命名 ChatGPT Work 听着像企业版，其实想对标 Cowork。

---

• 

  pEYTONwERNER

  —

  实测下来 Work 在 Web 和产出型任务上更宽，Sites 和生图是独有优势，20 美元 Plus 拿到的东西超出定价。但重度依赖 Microsoft 365 写操作的还是 Cowork 更成熟，OneDrive、SharePoint、Outlook 读写它更早做。我是 Web 活儿用 Work、本地文件流程用 Cowork，两个互补。

---

• 

  SRossSr

  —

  GPT-5.6 那 36 个变体配置表看得我头大，就一个滑块还要分 Max、Ultra 努力级别，社区都在喊把自动路由加回来。

---

• 

  CGonzales_66

  —

  看到好几个 Windows 用户在沙箱里反复撞墙、浪费 token，官方也承认开发资源长期偏向 Mac。我是 Mac 用户影响不大，但团队里用 Windows 的同事跑 Codex 任务老卡，只能开 --yolo 模式裸奔，挺无奈的。希望 5.6 真能改善跨平台体验。

---

• 

  EliKrauss

  —

  用 Sites 做了个发布日历给团队，底层表格一改它自动更新，比天天手动发周报强太多。

---

• 

  TheresaVasquez_2023

  —

  Ethan Mollick 说看不懂 Work 是什么、用它替代 Codex 得到什么失去什么，我也是。实测下来 Work 和 Codex 功能、历史记录完全一样，就是两套 UI 皮肤。如果你不怕 Git，其实没理由切到 Work 模式，纯聊天反而被边缘化成伪 IDE 了。OpenAI 这波有点为整合而整合。

---

• 

  蝴蝶_1

  —

  销售团队用 Work 把探索性对话转成 PoC，原来几周现在不到 24 小时，老板直接惊了。

---

• 

  GMartinez_66

  —

  上线第一天就冲了 Plus，Work 比想象中能打。

No reviews yet