Claude, ChatGPT Work, or Gemini Spark: pick by what the agent can reach
What each of the three AI work agents is allowed to touch, what stops it mid-task, which plan and country unlock it, and how to test one on a real task in an afternoon.
If your assistant could take one recurring job off your desk and hand it back finished, which would you give it to? In September 2026 all three of the big assistants made that offer for about $20 a month, and each went about it differently.
Anthropic merged Claude Cowork into the main Claude app on 16 September 2026, so handing over a report is now a mode inside an ordinary conversation, rolling out first to Pro and Max subscribers with Team and Free to follow. 1 OpenAI's ChatGPT Work has been rolling out since 9 July 2026, starting with Pro, Enterprise and Edu accounts and reaching Plus and Business afterwards. 2 Google's Gemini Spark launched on 19 May 2026 behind the AI Ultra subscription and now comes with the cheaper AI Pro plan as well. 34
Check the door before you compare feature lists, because the three doors are shaped differently. Claude's hand-off and scheduling need a paid plan, and the merged version is reaching Pro and Max subscribers first. 15 Work in the cloud needs a subscription at $20 a month or above, while the ChatGPT desktop app carries Chat, Work and Codex on every plan including Free, where Work is limited. 6 Gemini Spark asks for a Google AI Pro or Ultra subscription, an age of 18 or over, and a personal Google Account: work and school accounts are excluded, and it is unavailable across the European Economic Area, Nigeria, Switzerland and the United Kingdom. 7
The quick pick
| If the job looks like this | Start with | Why |
|---|---|---|
| A report, deck or spreadsheet that comes back on a schedule | Claude | Hand-off and scheduled tasks sit inside the paid Claude app, a cloud session keeps running after you close the laptop, and you can check progress from your phone. 145 |
| Long work that needs code, a live browser and files that persist between runs | ChatGPT Work | Work gets its own code execution with internet access, a full headless Chrome, a persistent workspace, sub-agents and scheduled prompts. 69 |
| Work that already lives in Gmail, Calendar, Drive, Docs and Sheets | Gemini Spark, where your account and country qualify | Spark connects natively to Google's apps and runs tasks in the background around the clock; those connections start switched off until you turn them on. 710 |
| A first trial | The assistant you already pay for | All three entry plans cost about the same, so the cheap experiment is the subscription you have already bought. 456 |
The comparison at a glance
| Dimension | Claude | ChatGPT Work | Gemini Spark |
|---|---|---|---|
| Where the work runs | Cloud sessions on Anthropic's servers, reaching your files, browser and apps through the desktop app. 3 | In the cloud on web and mobile; in the desktop app it can also run on your own machine with local files and apps. 69 | In Google's cloud, or in your own Chrome once you allow auto browse on that device. 7 |
| What it can reach | Folders and tools you choose, a browser built into the side panel, and whatever plugins and connectors you install. 34 | Code execution with internet access, headless Chrome, a filesystem that persists between sessions, GitHub repositories, and sites it can build and deploy. 9 | Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube and Maps, plus the sites you are signed into when it browses for you. 7 |
| What stops it | Asks before acting by default; an automatic mode screens each action before it runs; deleting a file always needs an explicit allow. Computer use has no sandbox between Claude and your screen. 13 | Local files and desktop apps need your permission. Cloud work runs in a container that can reach the open internet. 69 | Confirms before every browsing task, and some tasks still ask for confirmation. Sign-in steps during a remote-browser run wait for you to take over. 7 |
| Entry price | Pro at $17/month billed annually ($200 up front) or $20 monthly; Max from $100/month. 5 | Plus at $20/month; Pro from $100/month. 612 | Google AI Pro at $19.99/month; Ultra at $99.99/month. 4 |
What Claude feels like in use
Handing work to Claude now starts in the same conversation you would otherwise ask a question in. The separate Cowork tab is gone, and what it could do is available from any conversation: read folders you grant, open its own browser in a side panel, work in the background, and run a task on a schedule, with your skills, connectors and context already attached. 14 A cloud session reaches your computer only while the desktop app is open, and only for the folders you have connected there. 3
The permission model is what to read before you delegate anything. Claude asks before it acts unless you change that, and you can switch it to keep working and check in only when something needs a closer look. An automatic mode still reviews each action for safety before it runs, and Anthropic also offers a mode that skips approvals altogether, which its own safety page describes as having nothing check the agent's actions. 13 Permanently deleting a file always needs your explicit allow. 3
Simon Willison tried Cowork on its first day and traced where its permissions came from: the files it worked on were mounted in a containerised environment, which he later found was a custom Linux system booted through Apple's virtualisation framework. He still named prompt injection, where an instruction planted in a page the agent reads redirects its plan, as the risk that survives that sandbox. 8
What ChatGPT Work feels like in use
Work is two products wearing one name. In the cloud, on the web, on mobile or from the desktop app, it runs remotely and its chats sync across your devices. In the desktop app it can also run locally against your files and desktop apps, and that local mode is the reworked Codex. 69
The capabilities that separate Work from an ordinary chat are the reason to choose it. Work can execute code with internet access, drive a full headless Chrome instance, and keep a workspace that survives between sessions, so a task can clone a repository, install what it needs, and leave the result somewhere the next run will find it. 9 That makes Work the one to pick when the deliverable is something built, such as a script, a spreadsheet assembled from four regional exports, or a small internal site. 9
It is also the hardest of the three to predict. Simon Willison, after a month with it, called Work "extraordinarily confusing", noted that OpenAI's guidance describes what Work is for while its mechanics go undocumented, and said the system prompts and tool descriptions stay hidden. He also described how its abilities line up with what security people call the lethal trifecta: access to private data, exposure to untrusted content, and a route for sending information back out. 9
What Gemini Spark feels like in use
Spark is the one that keeps working when your hardware does not. It runs tasks and schedules in the background, and Google's description is that it carries on even with your phone and laptop switched off. 10 The connections that give it anything to work with are off until you switch them on, and then it reads across Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube and Maps. 7
Spark browses in two ways. On your own machine it uses the Chrome you are signed into, with all the access that implies, and it needs that device awake and running; if the device becomes unavailable, Spark may continue in a remote browser, where a sign-in prompt stops the task until you take over. With your permission it can use credentials saved in Password Manager. 7
PCMag's Ruben Circelli called Spark the best AI agent he had tested and still found the same reliability problems that affect its rivals. He also pointed out that much of what Spark demonstrates already exists in Gemini without the Ultra subscription, and that Spark was Ultra-only when he tested it. 11
Pick by the job it has to finish
A recurring deliverable. Choose Claude when the work repeats on a schedule and the output is a document, deck or spreadsheet you review each time. The hand-off, the schedule and the phone check-in are the product, and the permission default means the review step is built into the loop. 14
Work that has to be built. Choose ChatGPT Work when the task needs code, a live browser and files that outlive a single session. If your task finishes as a written answer, the code execution and headless Chrome are machinery you will pay for and leave idle. 9
Work already inside Google. Choose Spark when the task is written in Gmail, timed in Calendar, or sourced from Sheets and Docs. The same job in Claude or Work means wiring up connectors that Spark already has. 7
A managed laptop. When your employer owns the machine, the deciding question is administrative: Work's availability depends on your workspace settings and role-based access controls, and Spark does not accept work or school accounts at all. 67
The traps that change the choice
One mode removes the approvals entirely
Claude ships with approval prompts switched on, and the same product also offers a mode where nothing screens the agent's actions. Computer use is the sharper edge: when Claude takes over your screen it works through your apps and browser without the sandbox that surrounds its file work, and a link it clicks opens even in an app you never granted it. 13
Region and account type decide Spark before features do
Spark needs a personal Google Account and is unavailable across the European Economic Area, Nigeria, Switzerland and the United Kingdom. For readers in those places the agent is out of reach whatever it can do. 7
The tier above $20 is where the metered work sits
Each vendor sells a heavier plan, and each one is metered. Claude Max starts at $100 a month. ChatGPT Pro starts at $100 a month, with new sign-ups and upgrades to the $200 tier paused since 10 September 2026. Google's Ultra plan is $99.99 a month. 4512
A schedule you cannot stop
Spark can run a schedule while you are offline, and Google warns that you may not be able to stop it finishing an unintended action in that window. 7
Accountability arrives after the fact
A report published on 12 September 2026 alleges that an OpenAI agent swarm ran an attack on the RubyGems package registry in May. OpenAI said on 11 September that it was investigating and had not verified the specific claims, adding that its agents had used RubyGems to reach public information for benign tasks. Whatever the final finding, the episode is a reason to keep a person at the approval step on any machine that matters. 13
A one-afternoon test with the same material
Use one small packet you can check by hand: three documents, one spreadsheet, and a task whose right answer you already know. Run the whole thing with approval prompts switched on.
- Ask for one specific deliverable file and say what finished means.
- Count the approval prompts, and note which actions ran without asking.
- Compare the output against your own answer, and leave its own description of the work out of it.
- Ask it to name the file and section behind every number it used.
- Run the same task a second time, or on a schedule, and see whether the result matches.
- Switch to manual approval and confirm that the task still finishes.
Record four things: how many approvals it asked for, how many corrections the output needed, whether every number traced to a source you can open, and how long the task took. The agent that wins on those four is the one worth building a recurring workflow on.
Bottom line
Choose Claude when the recurring deliverable is the point and you want the task scheduled inside the app you already chat in. Choose ChatGPT Work when the work has to be built with code and a browser. Choose Gemini Spark when the task lives in Google's apps and your account and country let you in. Then let one real task, run with approvals on, settle it.
References
- 1
- 2
- 3
- 4Google AI plans
one.google.com
- 5Claude pricing
claude.com
- 6ChatGPT pricing
chatgpt.com
- 7Use Gemini Spark to manage your tasks & workflows in Gemini Apps
support.google.com
- 8First impressions of Claude Cowork, Anthropic's general agent
simonwillison.net
- 9Understanding ChatGPT Work
simonwillison.net
- 10Gemini Spark: Your 24/7 personal AI agent
gemini.google
- 11
- 12About ChatGPT Pro tiers
help.openai.com
- 13OpenAI agents attacked RubyGems back in May
simonwillison.net
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.