Claude Cowork評測:手機版全開放,但砍檔案風險未解 | Claude Cowork Review: Mobile Rollout Done, Risks Remain
By Kit 小克 | AI Tool Observer | 2026-08-22
🇹🇼 Claude Cowork評測:手機版全開放,但砍檔案風險未解
Claude Cowork從8月18日起正式對Anthropic所有付費方案(Pro、Team、Max)開放手機與網頁版,結束了先前只有Max用戶優先體驗、其他人只能乾等的階段。這代表現在不用整天守著電腦,也能在手機上啟動、監控、甚至核准一個多步驟的AI代辦任務。
Claude Cowork是什麼
不同於一問一答的Chat模式,Claude Cowork是Anthropic主打的AI Agent工作平台,能串接檔案、信箱、行事曆、瀏覽器等工具,自己拆解任務、跨多個步驟執行到完成,主打程式開發以外的一般商業與內容工作。
這次更新了什麼
- 跨裝置接續:桌面開任務,手機看進度,結果隨處可查。
- 背景執行:不需要裝置一直在線,排程任務可在伺服器端獨立跑完。
- 需要你拍板時會通知手機:Cowork卡在決策點時會推播請你確認,不用整天盯著。
- Chat與Cowork介面合一:桌面版和網頁版現在共用同一個首頁、同一個搜尋。
先別急著全權放手:真實用戶回報的風險
網路上已經有不少踩雷案例,用了才知道Claude Cowork不是萬能:
- 誤刪檔案:有用戶請Cowork「整理」雜亂資料夾,結果它把重要資料當成「雜物」直接刪除,一次損失11GB。
- Token燒得很快:因為Cowork會把多個檔案整份讀進上下文,有人一個下午除錯就燒掉50美元額度,還是反覆重讀同10個檔案造成的。
- 連接器不穩:Google Drive、Calendar整合時常認證失敗,或處理大資料夾時逾時。
- 沒有長期記憶:專案結束session就清空,除非把資訊重新丟進檔案,否則Cowork不記得先前的決策。
對手也到齊了:ChatGPT Work
OpenAI在7月推出ChatGPT Work迎戰,跑在GPT-5.6上,能接Slack、Teams、SharePoint、CRM等企業工具,主打用量與一般聊天分開計算,不會佔用日常對話額度,這點目前是Claude Cowork沒有的優勢。而Cowork則是在微軟365的深度整合上(尤其是寫入權限)領先一步。
小克怎麼看
手機版全面開放確實方便,適合拿來跑「幫我整理報告草稿」「幫我彙整這週信件重點」這類低風險任務。但只要涉及刪除、覆寫檔案的操作,權限一定要卡緊,先在測試資料夾試跑,別直接指向正式工作目錄。Claude Cowork還在快速迭代中,能力和穩定度都不是終點。
好不好用,試了才知道。
🇺🇸 Claude Cowork Review: Mobile Rollout Done, Risks Remain
Claude Cowork rolled out to mobile and web for all Anthropic paid plans, Pro, Team, and Max, on August 18, 2026, closing out a phased rollout that started with Max subscribers getting early access in July. That means you no longer need to be at your desktop to kick off, monitor, or approve a multi-step AI task.
What Claude Cowork Actually Is
Unlike a standard chat session, Claude Cowork is Anthropic's agentic work platform: it connects to files, email, calendar, and the browser, breaks a task into subtasks, and runs it end-to-end. It's pitched at general business and content work, not just coding.
What's New in This Update
- Cross-device continuity: start a task on desktop, check progress on your phone, retrieve results anywhere.
- Background processing: scheduled jobs run server-side with no device required to stay online.
- Mobile approvals: when Cowork hits a decision point, it pings your phone for input.
- One unified home: chat and Cowork now share the same sidebar and search across desktop and web.
Do Not Hand Over the Keys Yet: Real Risks Users Are Reporting
Early adopters have already hit real problems worth knowing before you trust Claude Cowork with anything important:
- Accidental deletion: one user asked Cowork to clean up a messy folder, it treated important files as clutter and deleted 11GB in one pass.
- Fast token burn: because Cowork loads multiple files into context at once, one user burned 50 dollars in credits in an afternoon debugging a circular dependency, with the same 10 files getting re-read repeatedly.
- Flaky connectors: Google Drive and Calendar integrations frequently fail to authenticate or time out on large folders.
- No persistent memory: once a project ends, its context ends with it, Cowork won't remember prior decisions unless you feed the same info back in.
The Competition Just Showed Up: ChatGPT Work
OpenAI launched ChatGPT Work in July to compete directly, running on GPT-5.6 with connectors for Slack, Teams, SharePoint, and CRM tools. Its usage is billed separately from regular chat conversations, an advantage Claude Cowork doesn't currently match. Cowork, meanwhile, still leads on Microsoft 365 depth, including write access added this July.
Kit's Take
The mobile rollout is genuinely convenient for low-risk tasks like drafting summaries or triaging your inbox. But anything involving deletion or overwriting files needs tight permission scoping, test in a throwaway folder before pointing Claude Cowork at your real working directory. This is still a fast-moving product, not a finished one.
好不好用,試了才知道。
Sources / 資料來源
- Claude Cowork on web and mobile: hand off work anywhere (Anthropic)
- Anthropic says Claude Cowork is now available on mobile for all paid plans (9to5Mac)
- ChatGPT Work vs Claude Cowork: Which one to choose in 2026? (Composio)
延伸閱讀 / Related Articles
- AI心智病毒評測:Anthropic證實Agent間會互相感染 | AI Mind Virus Review: Agents Can Infect Each Other
- AI自主駭客攻擊評測:全球首例AI Agent攻陷台灣政府 | AI Agent Cyberattack Review: First Autonomous Hack Hits Taiwan
- Anthropic IPO評測:2兆美元估值挑戰史上最大上市案 | Anthropic IPO Review: $2 Trillion Bid Beats SpaceX Record
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言