Open the Muse app, go to Settings, and look at the meter next to Manage subscription. Say it reads 73% used.

Seventy-three percent of what? A week of chatting? A month of errands? A pile of tokens with a name you have never seen on any other bill in your life?

Meta will not say. That silence is the whole pricing story, and it is the reason most people will buy the wrong plan.

A price list with the unit missing

Credit Meta for the part it does publish. Its subscription help page lists three tiers in plain numbers, and independent summaries agree on them word for word.

Free: $0, with a usage limit. Meta's help page does not publish the size of that limit. Mark Zuckerberg put it at up to 100 million Muse tokens a week in his launch post, which is a CEO statement, not a help page, and the difference matters when the bill arrives.

Power: $20 a month, 500 million Muse tokens a week.

Maximum: $100 a month, 3 billion Muse tokens a week.

Plans renew monthly on the date you subscribed. Meta lists no annual option. If you bought through the Apple App Store or Google Play, you cancel there, not in the app. You must be 18 or older, and Meta notes that benefits and availability vary by region and account, because the whole thing is still in limited testing in the US and Canada.

Simple enough to fit on a receipt. Now try to spend it. What does one Muse token buy? One message? One minute of browsing? One scheduled job? Meta has never published a conversion, and that gap turns every tier into a guess.

Run the math Meta skipped

Arithmetic is free, so run it yourself. A month holds about 4.33 weeks, which gives you monthly budgets of roughly 433 million tokens on free (using Zuckerberg's 100 million figure), 2.17 billion on Power, and 13 billion on Maximum.

Divide price by budget and you get about $9.20 per billion tokens on Power and about $7.70 per billion on Maximum. Read that twice, because it kills the obvious sales pitch. Paying five times as much does not buy tokens cheaper in any meaningful way. It buys headroom. You are paying $100 for permission to run six times as much stuff, not for a bargain on the stuff itself.

Notice the hole in the middle of the ladder. Nothing exists between $20 and $100, which means nothing exists between 500 million and 3 billion tokens a week. A user who genuinely needs 800 million a week has two choices: pay for 3 billion or wait for the meter to refresh. Somewhere between those two options, a lot of money gets left on the table or a lot of work gets left undone.

When you hit the limit, there is no top-up and no pay-per-use for the consumer agent. You upgrade, or you wait. That is a deliberate shape: a cliff, not a slope.

What a token buys, according to the only witness

Meta does not convert Muse tokens into messages, tasks, or hours of agent work. Longer tasks, larger files, and work in the agent's browser use more than a short chat, and that sentence is about as specific as any official source gets.

One reviewer's account is the best public evidence available, and it deserves both attention and a warning label. Testing the free tier, that reviewer found that building drawings, MIDI songs, small games, and animated websites used about 6% of a 100 million weekly allowance. Regular daily use over two weeks took about 35%. Job scans running every few hours ate about 31% of a week.

One account, one setup, not a study. Treat those numbers as a weather report from one town, not a climate model. Still, they point at the pattern that matters: background jobs burn far more than conversations. A chat costs a little. A scheduled scan that wakes up every few hours and browses the web costs a lot, every day, while you sleep.

Compare that with Manus, a rival agent maker, which sells credits at $20 a month for 4,000 and publishes example task costs so a buyer can translate credits into jobs. Meta could publish the same kind of table tomorrow. It has not, and the absence is doing real work for the company: a buyer who cannot translate the unit cannot comparison-shop it.

Measure your week before you pay for one

Forget the marketing numbers and use the one instrument Meta does give you. It sits in Settings, next to Manage subscription, and it shows the percentage of your allowance left. That meter is the only price signal in the whole system, so use it like one.

Step one: stay on free for one normal week. Not a light week, not a vacation week. A week with your real errands, your real scheduled jobs, your real late-night questions.

Step two: read the meter at the end of that week and write the number down.

Step three: apply a simple rule. Under about 60% in a normal week, stay free. Regularly over it, buy Power. Only if Power itself empties before the week ends should Maximum even enter the conversation.

Step four: count your scheduled jobs before you count your messages. If your meter is high, the culprit is almost certainly a background task, a price watch, a job scan, a daily digest, running on a schedule you forgot you set. Kill or thin the schedule first. Paying $100 a month to fund jobs you no longer want is the most expensive way to avoid a five-minute audit.

One more number for the road. An invite code adds 1 billion tokens once, to you and to the person who posted it, which works out to roughly $9 of allowance at Power's rate, or about ten weeks of free-tier use at Zuckerberg's 100 million figure. Code pages exist because each redemption also pays the poster. Never pay for a code, and never treat a one-time bonus as a plan.

Who pays for free

Skepticism belongs here, because free is never free, it is just billed somewhere you are not looking. At Connect on September 23, Zuckerberg said Meta is making Muse free for a huge number of tokens, with the expectation that over time the company will profit by taking a small fee from transactions.

Look at what got wired up alongside that sentence: Shopify's full catalog with Shop Pay, PayPal, Stripe, retail connectors for Best Buy, Gap, Sephora, Walmart, and Wayfair, travel through Expedia. Your agent is being built into a checkout lane, and the company that owns the lane takes a cut of what passes through it.

That creates a question no approval screen answers yet. When your agent picks between two flights, two shirts, two hotels, and one of them pays Meta a fee, does the screen tell you which one? Approvals gate whether the agent spends, not whether the agent's recommendation is neutral. Before you hand purchasing over, ask who gets paid when your agent chooses the more expensive option, and keep asking until the answer appears in the product, not in a keynote.

Three price lists, same words, different bills

Last trap, and it is an expensive one. "Muse" is three products with three price lists, and two of them reuse the words Power and Maximum with completely different numbers.

Muse Code, the developer coding agent for the terminal, charges Everyday $5, High $15, Power $50, and Maximum $200 a month. A Muse Code plan covers only its command-line key. It does not raise your consumer Muse allowance by one token.

Meta's Model API bills developers pay-as-you-go per token, per image, or per hour of audio. Again, a separate bill, a separate product.

Check which product you are paying for before you upgrade, because "I bought Power" can mean $20 or $50 depending on which Muse you meant, and support will not find that funny.

The verdict

Most people should pay nothing. A normal week of errands, reminders, and questions fits inside the free allowance with room to spare, and the meter will tell you so if you let it run a week.

Pay $20 if your meter regularly crosses 60% in a normal week, or if you run daily scheduled jobs that browse, scan, or summarize while you are away. That is the tier for people whose agent works a shift, not a cameo.

Pay $100 only if Power itself runs dry, which means your agent is doing heavy, continuous, background work, the kind small businesses run across connected systems all day. If you are not sure whether that describes you, it does not.

Whatever you pay, remember what you are actually buying. Not tokens, not tasks, not hours. Headroom on a meter whose unit nobody will define, inside a product whose free tier is funded by a future cut of your purchases. Measure first, pay second, and audit the scheduled jobs before you upgrade the plan. The meter is honest even when the marketing is not. Use it.

What this does not prove

Honest accounting, as always. No firsthand meter test sits behind this article, and no Meta confirmation of the free tier's size exists in Meta's help pages; the 100 million figure is Zuckerberg's launch claim, reported by CNET and Gizmodo, and it could change without notice. Burn rates from one reviewer's account are anecdotes, not measurements, and the monthly budgets and per-billion prices are my arithmetic on Meta's published weekly allowances, not Meta figures. Invite-code values use the same arithmetic. Availability is limited to the US and Canada, and launch coverage disagreed on whether signup requires a payment card even on the free tier, so treat that detail as unresolved rather than settled. What stands without qualification is structural: a pricing unit with no published conversion cannot be comparison-shopped, a ladder with no middle rung forces overbuying, and a free tier funded by transaction fees deserves a question about neutrality before it gets your wallet.

Sources: Meta, About Muse subscriptions (subscription help page), via ToolColumn, ToolColumn limits guide, Muse Directory, and TrendInTech; Mark Zuckerberg launch post via Tech Insider; Zuckerberg at Meta Connect, September 23, 2026, via The Signal; Roborhythms free-tier usage via iTechnically Kan; Manus credit pricing via ToolColumn; invite-code bonus via Vandata Team and Top5Apps; Muse Code and Model API pricing via Muse AI Alternative.