Meta Muse vs ChatGPT vs Claude vs Gemini: Who Does the Work

Four companies now sell me the same sentence: give the agent a goal and walk away. Meta joined on Monday with Muse. OpenAI has had ChatGPT Work since July, Anthropic has had Claude Cowork since April, Google has run Gemini Spark since I/O in May. Same sentence, four very different products, and I cannot get the newest one. Here is how they line up on the things that decide whether an agent is useful: what it can touch, what it costs, where it runs, and who you have to trust.

Four Agents, Four Different Ideas of Personal

Start with what each company thinks a personal agent is for, because the products only make sense through that lens.

Meta's Muse is for errands. The launch list is emails, travel, bills, forms, party invites, grocery lists from recipe reels, and checking out at stores. It lives in a dedicated virtual machine in Meta's cloud, keeps working after you close the app, remembers your preferences, and pings you when it needs a yes. It talks to you through its own app, muse.ai and WhatsApp. I went through the whole launch, including the security architecture and the internal test leaks, in Meta Muse Is Here, so this piece assumes that.

OpenAI's ChatGPT Work, launched July 9, is for deliverables. You describe an outcome, it pulls context from Slack, Gmail, Drive, Salesforce and the rest of its plugin directory, works for hours if it has to, and hands back a spreadsheet, a deck, a report or a small web app. It can run on a schedule. It sits as a third mode next to Chat and Codex, and OpenAI's help center now sends people who ask for the old agent mode to Work, so this is the agent in ChatGPT. It shipped on GPT-5.6. GPT-6 Astra has been on Plus and up since September 3.

Anthropic's Claude Cowork is for your files. It went GA on April 9 as the third tab in the desktop app, reads the folders you grant, works across your applications, and since late August opens its own browser in a side panel when a task needs a website. Routines handle scheduled work, Live Artifacts re-run their code when you reopen them, and as of this month sessions can run remotely from web and mobile, in beta, with state synced to your account.

Google's Gemini Spark is for your Google account. Announced at I/O on May 19, it runs around the clock on Google Cloud VMs, watches your Gmail, Docs and Calendar, takes pre-authorized actions, and surfaces results when they matter. You can email it at a dedicated Gmail address. It browses through Chrome and reports progress on Android through the new Halo system. It moved to Gemini 3.7 Flash on August 13 and is open to AI Pro and Ultra subscribers in more than 160 countries.

Errands, deliverables, files, your inbox. That is the whole comparison in one line. The rest is detail, and the detail is where the money goes.

Price, Free Tier, and Who Can Sign Up

Muse is the only one with an agent on the free tier. 100 million tokens a week, per Zuckerberg's Threads post, with a card on file. Gizmodo burned through an evening of game and website generation and used 11 percent of it. The others give you a chatbot for free and sell you the agent at 20 USD.

At 20 USD the four are interchangeable on price, and that is not an accident. Meta named its tiers Power and Maximum and priced them at exactly the ChatGPT and Claude numbers. Google cut AI Ultra from 250 USD to a new 100 USD tier plus a 200 USD tier at I/O, so its ladder matches too.

The real difference is who can sign up. Muse is US only and 18 plus. Spark launched US only for Ultra, then opened to Pro and 160 plus countries over the summer. Cowork comes with every paid Claude plan worldwide. Work comes with Plus, Pro, Business and Enterprise wherever ChatGPT is sold. If you sit in Europe, the newest agent on this list is the one you cannot have.

One more pricing note. Claude Pro includes Fable 5.1, Anthropic's top model, at up to 50 percent of your weekly limit, and the Chat, Code and Cowork tabs all draw from the same pool; which tier earns that price I worked through in Claude Fable 5.1 vs Opus 5. ChatGPT caps Astra messages per week on Pro. Meta has not published a per-tier token number at all.

Where They Run and What They Can Touch

The architecture decides what the agent can do, and what can go wrong.

Muse and Spark are cloud-native. Each user gets a VM on the provider's infrastructure, so the agent works while your laptop is closed and never sees your local disk. Muse's VM carries its own browser and a shell it can run code in. Spark browses through Chrome and speaks MCP for outside services. Neither can touch a file on your Mac.

Cowork is the opposite. It started as a desktop product with access to the folders you approve, which is why it is the one I use for actual work on actual repositories. The browser lives in a side panel on your machine, which I covered when it shipped in Claude Cowork Now Opens Its Own Browser. Remote sessions are new and in beta. If your job is "clean up this folder of client PDFs and draft the invoices," Cowork is the only one of the four that can see the folder.

Work sits in between: cloud-hosted, but aimed at your company's tools rather than your personal life. Connected apps, scheduled runs, long tasks, finished documents. It is not trying to book your dentist.

Then there is money, which is where Muse stands alone. It checks out through Link by Stripe with single-use card numbers bound to one merchant and one amount. Meta says it is the first agent covered by Link's purchase protections. OpenAI tried this. Instant Checkout launched inside ChatGPT on February 16 with Etsy sellers and a promise of a million Shopify merchants, and by March OpenAI had shelved it to "focus our efforts on product discovery." Anthropic has never put a payment rail in Cowork. Google has not announced one for Spark. If "does the agent spend money" is your test, Muse is the only yes today.

Memory is the softer edge. Muse learns preferences across conversations and will suggest things unprompted. Spark builds personal context from your Google data by design. Claude has memory across the app, though Cowork's context is scoped to the workspace you are in. Work leans on your connected company data rather than on you.

Trust Models Side by Side

Every one of these products asks you to hand over an inbox. The difference is how they justify it.

Meta's answer is the most engineered. A separate Sentinel process approves every network call, credentials are swapped in at the boundary so the model never holds a real token, approvals happen in the app UI rather than in the chat where an injection could fake them, and the browser sees an accessibility tree, not the DOM. It is also the answer with the biggest asterisk. Meta's own document says the VM "does not prevent Meta from accessing data when necessary to support, secure or operate the service," a Meta VP confirmed to Wired that access is technically possible, and Reuters saw internal test posts describing exposed iCloud photos. The Confidential VM that would lock Meta out is due "later this year."

Anthropic's answer is permission at the edge of your own machine. Cowork asks before it touches a folder, and the browser runs where you can watch it. The model side is Fable 5.1 with Anthropic's safeguards, which I looked at in Claude Fable 5.1 Refuses Less. The trade is that the agent works on your hardware, so a mistake lands on your disk, not in a sandbox.

OpenAI's answer for Work is narrow scope: company tools in, documents out, and no payment rail since March. Google's answer is that Spark stays inside the account you already gave Google, acting "under your direction." Forbes ran a piece in May under the headline "Google Announced Gemini Spark, But Left Out An Uncomfortable Warning," which tells you how the keynote handled risk.

Then there is the wildcard none of the four want to talk about. Instinct, an invite-only agent from a company founded in April, went from stealth to a 2.5 billion USD valuation within days in late August by doing roughly what Muse does with fewer guardrails, and testers found that disconnecting their email did not delete the copied data. Meta will point at Instinct and say it built the sandbox. Fair. It is still Meta holding the sandbox.

Bottom Line

My routing, for what it is worth.

Files and code: Cowork, because it is the only one that can see them. Company documents from connected tools: Work, if you already live in ChatGPT. Inbox babysitting inside Google: Spark, and at 19.99 USD it is the cheapest way in. Errands with a card attached: Muse, once it reaches Europe and once the Confidential VM ships. Not before.

What I keep noticing is that four companies converged on 20 USD and diverged on everything else. That reads like a placeholder price while they work out what an agent is worth. Meta's free 100 million tokens a week is the first real move away from it, and I expect an answer from at least one of the others before the year is out.

If you want a setup that keeps Claude's agent on a leash while it works in your repositories, that is what Claude Blueprint packages: the hooks and permission rules I run in every project. The rest of my Claude Code and Anthropic coverage is collected on the Claude Code topic page.

This article contains affiliate links. If you sign up through them, I may earn a small commission at no extra cost to you. (Ad)