Muse vs Work vs Cowork vs Spark: What Each Agent Touches

Four companies now sell the same promise: give an agent a goal and walk away. Only one of them can open a file on your laptop, and only one can spend your money.

Meta's Muse landed on September 8, joining OpenAI's ChatGPT Work, Anthropic's Claude Cowork and Google's Gemini Spark. The marketing converges. The capabilities do not. Here is what separates them on the four questions that decide whether you can use one at work: what it can reach, where it runs, whether you can buy it, and what you are trusting.

What each one is actually for

A comparison published this week frames the split in one line — errands, deliverables, files, your inbox — and the launch details behind it hold up across the outlets covering Muse.

Muse is for errands. The Verge reports it can open a browser, fill out forms, and negotiate on users' behalf, keep working in the background after you close the app, and come back when it needs approval for something like a purchase. TechCrunch describes it as wanting access to users' email, calendars, payments, health services and more.

ChatGPT Work, launched July 9, is for deliverables — you describe an outcome, it pulls context from connected tools and hands back a spreadsheet, a deck, a report or a small web app, and it can run on a schedule.

Claude Cowork, GA since April 9, is for your files. It reads the folders you grant, and since late August it opens its own browser in a side panel when a task needs a website.

Gemini Spark, announced at I/O on May 19, is for your Google account — it watches Gmail, Docs and Calendar and takes pre-authorized actions.

Where they run decides what they can reach

This is the distinction that actually gates real work, and it splits the field two-and-two.

Agent	Runs on	Can read local files	Can pay	Available
Meta Muse	Your own VM in Meta's cloud	No	Yes	US only, 18+
ChatGPT Work	OpenAI's cloud	No	No	Wherever ChatGPT is sold
Claude Cowork	Your desktop; cloud sessions in beta	Yes	No	Every paid Claude plan, worldwide
Gemini Spark	Google Cloud VM	No	Not announced	160+ countries

Muse and Spark are cloud-native: each user gets a VM on the provider's infrastructure, so the agent keeps working while your laptop is closed and never sees your local disk. As the comparison puts it, neither can touch a file on your Mac.

Cowork is the opposite trade. It started as a desktop product with access to folders you approve, and the browser runs on your machine where you can watch it. The cost of that reach is that a mistake lands on your disk rather than in someone's sandbox.

The practical test is blunt. If the job is "read this folder of client PDFs and draft the invoices," Cowork is the only one of the four that can see the folder. If the job needs to run for six hours while your laptop is shut, it is one of the other three.

Only Muse spends money — and OpenAI already tried

Muse checks out through Link by Stripe using single-use card numbers bound to one merchant and one amount, per the comparison. That is the one capability with no equivalent elsewhere on this list today.

The history matters more than the feature. OpenAI shipped Instant Checkout inside ChatGPT on February 16 with Etsy sellers and a promise of a million Shopify merchants, then shelved it by March to "focus our efforts on product discovery." Anthropic has never put a payment rail in Cowork. Google has not announced one for Spark.

Read that as a signal about difficulty, not appetite. Three of four either abandoned agent payments or never shipped them, which is the context to hold when evaluating Muse's version.

Pricing: the one place the sources disagree

Be careful here, because the numbers circulating do not all come from Meta.

The comparison lists all four at a $20 entry price, top tiers at $100 or $200, and claims Muse is the only one with a free agent tier at 100 million tokens a week with a card on file, attributing that figure to a post by Zuckerberg on Threads. It also states Google cut AI Ultra from $250 to a $100 tier plus a $200 tier at I/O.

The Verge's launch coverage says something materially different about Muse: the assistant will be free for most users, with paid subscriptions "for people who want to do more," and that Meta did not specify the nature and costs of these paid plans, nor outline what limits there are for free users.

Treat the $20/$100 Muse tiers and the weekly token allowance as one writer's reporting, not as announced pricing. If your decision turns on what Muse costs, that number is not yet public.

The trust question, stated plainly

Every one of these products asks for an inbox. Meta's answer is the most engineered of the four, per the comparison: 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 chat where an injection could fake them, and the browser sees an accessibility tree rather than the DOM.

It also carries the largest caveat. The comparison cites Meta's own document saying the VM "does not prevent Meta from accessing data when necessary to support, secure or operate the service," reports that a Meta VP confirmed to Wired that access is technically possible, and says Reuters saw internal test posts describing exposed iCloud photos. The Confidential VM that would lock Meta out is described as due "later this year."

Anthropic's answer is narrower and easier to reason about: permission at the edge of your own machine, with the browser running where you can watch it. OpenAI's answer for Work is limited scope — company tools in, documents out, no payment rail since March.

How to choose today

Match the agent to the constraint, not to the demo:

• Local files or repositories — Cowork, because it is the only one that can see them.
• Long unattended runs — Work or Spark, both cloud-hosted with scheduling.
• Outside the US — Muse is not an option at all; Cowork and Work are the broadest.
• Agent-initiated purchases — Muse is the only yes, and the pricing is still unannounced.

What to watch next: whether Meta publishes actual tier pricing, and whether the Confidential VM ships before enterprises are asked to connect real inboxes. Both are open as of this week.