StartClaude Scheduled Tasks Finally Run in the Cloud
By Addy · August 11, 2026 · Editorial standards
On July 13, Anthropic moved Claude's Scheduled Tasks off your laptop and onto its own servers. Set a task for 6 a.m., close the lid, go to sleep, and Claude keeps going, no device online required. That reads like a small infrastructure change, the kind of release-notes line most people scroll past. It closes the one specific, repeatedly documented complaint against the most capable of the three major scheduled-agent implementations, and it marks a real shift in what these products are actually for.
Every major AI assistant has spent the past eighteen months quietly pursuing the same idea: stop being something you open and start being something that's already working when you check back in. That's a bigger claim than a feature list suggests. A chatbot that only responds when spoken to is fundamentally a different kind of product from one that acts on a schedule, touches your real accounts, and delivers a finished result without being asked twice. ChatGPT, Gemini, and Claude all reached for that idea inside the same window of time, under three different names, and arrived at three genuinely different implementations. The gap between them is the actual story, and it's worth walking through in order, because the names converged long before the substance did.
The Reminder App Era
OpenAI shipped ChatGPT Tasks first, in beta in January 2025. For a long stretch, it did roughly what a reminder app does: a daily news briefing, a nudge about something you'd forgotten, delivered on a schedule instead of on request. Useful, but narrow enough that one early reviewer described it as exactly that, a glorified reminder app, and asked why they'd use it over the to-do list app they already had. The honest answer at the time was that they mostly wouldn't, not for anything more demanding than a recurring summary.
Google followed that November with Gemini's Scheduled Actions, matching the basic concept: recurring prompts that run without the app open, gated behind a $20 monthly plan and capped at ten scheduled actions active at once. The feature came from the company with the deepest native access to Gmail, Calendar, Docs, and Tasks of any AI assistant on the market, which made the early reviews more pointed than they might have been for a less Google-native product. Reviewers still found the implementation thin the moment a workflow needed to touch anything outside Google's own walls. Half-baked was the specific word used by at least one reviewer who otherwise liked Google's broader AI push.
Claude Arrives Late, and Builds It Deeper
Claude's version didn't show up until February 2026, built inside Cowork, Anthropic's file-and-tool-connected workspace, and it launched desktop only, paid plans only. What it did differently was structural: every scheduled run got a full Cowork session rather than a single canned prompt, meaning real access to your files, your connected tools through Claude Connectors or MCP servers, your skills, and your plugins, identical to what a manually triggered session would have. A scheduled task wasn't a lightweight cron job bolted onto a chatbot. It was the same product you'd use by hand, just triggered by a clock instead of a click.
By March, Claude actually had three distinct flavors of this running in parallel: Cowork scheduled tasks for the general knowledge-work case, Claude Code cloud scheduled tasks for engineering workflows, and session-scoped tasks inside the Claude Code CLI for anything that only needed to persist for the length of a single working session. That's more surface area than either competitor had built, and independent reviews that month argued it was the version that had actually learned from where the other two fell short.
The same reviews flagged the obvious tradeoff just as plainly: it only worked while your computer was on and the Claude Desktop app was open. Close the laptop mid-run and the task stopped with it, full stop, no partial credit. For a feature explicitly pitched as "set it and forget it," that was a real, structural contradiction, not a minor rough edge.
How the Three Actually Compare
| ChatGPT Tasks | Gemini Scheduled Actions | Claude Scheduled Tasks | |
|---|---|---|---|
| Launched | January 2025 (beta) | November 2025 | February 2026 |
| Required tier | ChatGPT Plus or higher | $20/month plan | Pro, Max, Team, or Enterprise |
| Ran without device online | Cloud-based from launch | Cloud-based from launch | Only added July 13, 2026 |
| Access per run | Connected apps via Tasks | Primarily Google Workspace apps | Full session: files, connectors, skills, plugins |
| Cross-tool reach outside home ecosystem | Moderate | Documented as thin | Broadest, by design |
The pattern in that table is the whole story in miniature. ChatGPT and Gemini were built cloud-first because they were built chat-first, so device independence was never something they had to earn. Claude's device dependency existed precisely because Cowork's actual advantage, deep local file and tool access, was implemented as a desktop feature first. Fixing it wasn't a small patch. It meant re-architecting where Cowork itself runs.
The Names Converged Long Before the Capability Did
By June 2026, a side-by-side terminology comparison across all three products described their scheduling features as near drop-in replacements for one another, Tasks, Scheduled Actions, Scheduled Tasks, three names for what had become, on paper, the same idea. That same comparison noted the pattern extends well beyond scheduling: ChatGPT and Gemini both call their editable side-pane Canvas, Claude calls it Artifacts; ChatGPT and Gemini both call deep multi-step research Deep Research, Claude just calls it Research. The entire category had converged on a shared vocabulary while the underlying products kept diverging underneath it.
That same month, ChatGPT Tasks got a dedicated mobile management page, letting Android users see and edit every active task in one place instead of digging back through chat history to find one. It was a real usability upgrade, and also a sign of how much of the competition by mid-2026 was happening at the level of interface polish rather than underlying capability.
The naming convergence didn't mean the experience had converged with it. Hands-on testing published in July, after all three had been treating their scheduling features as mature for months, still described Gemini's actual cross-tool orchestration as disjointed. On paper, Gemini could read your email, pull notes from Keep, cross-reference NotebookLM, and turn the result into a shared Google Doc. In practice, the reviewer's conclusion was that it worked, but felt nowhere near as streamlined as the pitch suggested, and this was someone testing inside Google's own ecosystem, the one environment Gemini is supposed to have a structural home-field advantage in.
Closing the One Gap That Was Actually Named
Anthropic's July 13 update addresses the specific, narrow thing critics had been naming since March. Cowork itself moved off individual machines and onto Anthropic's own infrastructure, and expanded at the same time from desktop-only to a beta rolling out on web and mobile, starting with the Max plan. Scheduled tasks and their finished outputs now follow you across devices instead of living on whichever machine happened to be running when the task was created. Anthropic's own example is specific enough to be worth repeating exactly: schedule a 6 a.m. client-prep task, and Claude works through the relevant email threads, transcripts, and recent news, builds the briefing document, and leaves a follow-up email drafted, not sent, waiting for a human to actually approve it before it goes out.
That last detail matters more than it might look like on first read.
The Part Nobody's Fully Solved: How Much Should Run Unsupervised
Every one of these products is, by design, now capable of taking real action inside a person's actual accounts through agentic workflows on a recurring, unattended basis. That's the entire selling point, and it's also the exact shape of the question this publication has spent a lot of this year on: what happens when an autonomous system has standing access to something real, and nobody's watching it in the moment it acts.
Claude's choice to leave that example follow-up email drafted but unsent isn't an incidental design detail. It's a specific, deliberate line drawn between "gather information and prepare work" and "take an action with real-world consequences," and it puts the second category behind a human review step even inside a feature explicitly marketed as unattended automation. Anthropic's own numbers give some sense of the stakes involved in getting that line right: more than 90% of actual Cowork usage isn't software development at all. It's everyday business operations and content creation, the kind of work least likely to be reviewed carefully precisely because it's routine.
Neither ChatGPT's nor Gemini's public documentation makes an equivalent distinction as explicit as Claude's does in its own materials, which isn't the same as saying they lack any safeguards, only that the evidence available doesn't let this piece make a fair three-way claim on that specific point. It's the detail worth watching closest as all three companies keep expanding what their scheduled agents are trusted to do without anyone in the loop.
What This Is Actually Part Of
There's a broader business signal sitting next to all of this, worth citing with the right amount of caution attached. Appfigures tracked Claude passing ChatGPT in daily US download counts starting February 28, 2026, and holding that lead since, 149,000 downloads to ChatGPT's 124,000 by early March. That's a real, independently tracked number. It is not, on its own, proof that scheduling specifically drove it, correlation without a clean causal claim, and plenty else changed across both products in that window.
What is clear is what this feature category is actually being used for once people have it. Not autonomous coding, not novel demos, mostly the unglamorous work that was the entire pitch to begin with, reconciling documents, drafting reports, watching a folder for changes, closing out an inbox before dinner. Eighteen months, three companies, and the same four words on every landing page: describe it once, forget about it. Only one of them has spent that time actually closing the gaps reviewers kept naming instead of just matching the feature list, and even that one hasn't finished the harder question sitting underneath the convenience.
Previously on TheQuery: