2026.09.22 | AC | Agents | Meta Muse Unit Economics

摘要

1. What is Muse's 100M weekly free token allotment actually worth under Meta's two API price tiers；
2. Can subscription revenue cover free-tier compute costs, and should Muse be valued as SaaS；
3. Does the free allowance buy user acquisition or training data；
4. Should the opportunity cost be measured at book depreciation or external rental prices；
5. How does the subsidy affect R&D cash flow, compute leasing, and consumer AI subscription pricing.

> Muse's free token allowance is, in substance, a data purchase. Its users were already inside Meta's ecosystem, and Muse conversations are walled off from the ad stack, so the only thing the give-away can buy back is training data. Marked at Meta's own external prices, each free weekly active user costs roughly $13.7 a month, and subscriptions recover only a small slice of that. The bill therefore lands in R&D and cash flow, while the advertising business, protected by the data firewall, is not directly exposed.

On September 8, Meta launched Muse, a personal agent that hands every user 100 million free tokens a week. Six days earlier the company had priced developer access to the same model generation: the standard tier charges $1.25 per million input tokens, while developers who let Meta train on their API traffic pay just $0.10 on the contributor tier. The same token carries a 12.5x spread across two rate cards — and drops to zero by the time it reaches consumers. The market reads the free allowance as customer-acquisition spend for the ad business. That reading does not hold. Which ledger absorbs the gap between these three prices is what determines how Muse should be valued.

I. A Free Tier Worth $18 to $339 a Month, and Only Bots Can Spend It

Converted at Meta's own two API tiers, 100 million tokens a week are worth roughly $18 to $339 a month. Meta's developer documentation prices Muse Spark 1.3, released September 2, at $0.15, $1.25 and $4.25 per million tokens for cached reads, fresh input and output on the standard tier, versus $0.002, $0.10 and $0.20 on the contributor tier. Agents re-read pages and context constantly, so cached reads dominate the mix; this report weights cached reads, fresh input and output 7:2:1, giving blended rates of roughly $0.78 and $0.04 per million tokens. At about 435 million tokens a month, the free allowance is worth some $339 at standard rates and some $18 at contributor rates — a spread of roughly 19x. That gap is the price Meta concedes in return for training rights, a point Section III comes back to.

Muse's “tokens” cannot be mapped one-for-one onto API-billed tokens; 100 million looks more like a marketing unit. The paid tiers give the game away. Power is reported to carry 500 million tokens a week, a figure Meta's official pricing page never states. On that basis, a fully consumed Power allowance implies about $0.009 per million tokens, roughly 1.2% of the blended standard rate — and only 22% of the blended contributor rate, which is itself the floor price Meta accepts in exchange for data. If both token definitions were the same, Meta would be selling compute at about one-fifth of its own data-for-price floor, which makes no commercial sense. Muse tokens and API tokens are therefore almost certainly metered on different bases.

Whichever meter applies, this allowance binds round-the-clock automation; human use will rarely come near the ceiling. Burning 100 million tokens means an agent running every hour of the week at roughly 600,000 tokens an hour. A Gizmodo reporter set out to squander compute on launch day — three small games, a website, 50 music tracks and a simulated Mac running inside the browser — and by Muse's own counter still consumed only about 11% of the weekly allowance. Only persistent, scheduled, multi-threaded workloads can exhaust it, and “working for you around the clock” is exactly how Muse is sold. There is precedent: in July 2025 Anthropic imposed weekly usage caps on Claude subscriptions, aimed squarely at users running Claude Code in the background non-stop; fewer than 5% of subscribers were affected. The cost of a free tier, in other words, concentrates in a small minority. Muse's unit economics therefore hinge on two variables Meta has not disclosed: how utilisation is distributed, and which rate card the cost is booked at. The next section takes the first.

II. Subscriptions Cover a Fraction of the Free Tier, So Muse Is Not SaaS

At contributor rates, the $20 Power tier breaks even at roughly 110 million tokens a week — essentially where the free allowance ends. Break-even usage is the monthly fee divided by the blended rate. On contributor pricing, the $20 Power tier and the $100 Maximum tier break even at about 110 million and 560 million tokens a week; on standard pricing, break-even is only about 5.9 million and 29.5 million. Fold in virtual-machine and browser time — at E2B's published $0.109 per hour and an assumed 1.5 million tokens processed per hour — and Power's contributor-rate break-even falls to about 40 million tokens a week, roughly 40% of the free cap. The catch is that users have a reason to pay only once the free allowance stops being enough. The moment usage clears that cap, tokens alone eat about 90% of the $20 fee even at Meta's cheapest rate; add VM time and the fee sits below cost. A paying Power user, put plainly, is inside Meta's loss zone from the first month. Meta's own framing agrees: subscriptions exist to defray heavy users' compute, and the overwhelming majority of users are expected to stay on the free tier.

Compute cost per free weekly active user works out at roughly $1.8 to $13.7 a month, with the 1% automation cohort absorbing about 27% of the total. This report splits free users three ways: 90% light users at about 2 million tokens a week, 9% core users at about 10 million, and 1% automation users maxing out at 100 million. At contributor rates the three cohorts cost about $1.0, $5.0 and $49.5 a month, a weighted $1.8; at standard rates, about $7.4, $37.0 and $370.5, a weighted $13.7. The cost mix also differs between the two bases. At standard rates, VMs and browsers make up about 9% of per-user cost, close to the roughly 8% CPU share of agent marginal cost estimated in our September 7 GPT-6 Astra report; at contributor rates that share climbs to about 64%, because VMs bill by the hour irrespective of token prices. As Meta's internal token costs fall, general-purpose compute takes up a larger share of the bill, so the free tier's total cost will not decline in proportion to token prices.

On industry-standard conversion rates, subscriptions can cross-subsidise the free tier only if paying users barely touch compute. RevenueCat's 2026 State of Subscription Apps report puts median day-35 download-to-paid conversion at 2.1% under freemium and 10.7% behind a hard paywall, with annual renewal for AI apps at 21.1%, below non-AI apps' 30.7%. Assuming a 9:1 Power-to-Maximum mix, conversion rates of 2.1%, 5% and 10.7% translate into roughly $0.59, $1.40 and $3.00 of subscription revenue per weekly active user. Against standard-rate costs that covers only 4% to 22% of the free tier; against contributor rates, coverage runs from 32% to 163% — but only if subscribers consume almost no compute themselves. Subscribers are, by construction, the users for whom the free allowance is too small, so their usage should cluster near the cap. At contributor rates including VM time, each Power user costs about $49.5 of compute a month, at which point the subscription's contribution to the free tier turns negative. Growing the paid base thus widens the loss when measured at external selling prices. On either basis, Muse subscriptions cannot carry a SaaS label. And if subscriptions cannot recover the cost, the reason Meta bears it must lie outside subscriptions altogether.

III. Muse's Users Were Already Meta's; the Allowance Buys Training Data

Treating Muse as an acquisition subsidy for advertising requires two conditions, and neither has evidence behind it today. The first is that Muse brings in new users. Apptopia estimates that more than 95% of Muse users also use Facebook and 63% also use Instagram (TechCrunch, September 21), leaving very little genuine new reach. The second is that Muse's data can reach the ad system. At launch Meta committed that Muse conversations and VM data stay out of advertising; yet under Meta's announcement of October 1, 2025, Meta AI chat interactions have fed ad personalisation since December 16 of that year, the EU, the UK and South Korea excepted. One company running opposite data policies on two AI products indicates the separation is deliberate. As an earlier report noted, Google's restricted-data policy similarly bars Gmail and Calendar data from advertising use; we do not revisit that here. With neither new users nor ad data available as a return channel, the payback must land somewhere else.

Meta has already put a price tag on usage data, and Muse's free tier is the consumer edition of the contributor tier. The contributor tier trades discounts of about 92% on input and 95% on output for the right to train on developer traffic. Muse likewise defaults to using consumer interactions for model training, with an opt-out; conversations from users who do not opt out are used to improve models once key personally identifying information is stripped. The pricing logic is identical: the developer tier buys data with a discount of more than 90%, while the consumer tier pushes the discount all the way to 100%. The gap between standard and contributor pricing in Section I is, in short, the price Meta pays for training rights — Muse simply moves that payment to the consumer end.

What Meta wants to buy is precisely the class of data Muse Spark is weakest at producing. On Meta's self-reported September 2 numbers, Muse Spark 1.3 leads on coding and long-context tests but scores 66.9% on OSWorld 2.0, the benchmark for completing tasks on a computer — below Claude Opus 5's 68.3% and about 5.7 points behind the 72.6% OpenAI reports for GPT-6 Astra. We understand Meta's training signal has long skewed toward social contexts and lacks real task trajectories. Our September 9 report argued that capabilities which can be auto-scored get internalised into model weights fast. Muse converts genuine consumer requests into scorable samples carrying pass-or-fail outcomes, which maps directly onto that weakness. The data frontier is clearly drawn as well: Google explicitly forbids using Workspace API data to train non-personalised models, so what Meta can legitimately accumulate is action trajectories on the open web and on its own platforms. There is precedent inside the company: the $14.3 billion Meta paid for 49% of Scale AI was, in substance, a purchase of training data. Muse's free tier extends the same thinking to consumers.

Transaction commissions are the other plausible recovery channel, but for now they carry only option value. Meta's stated monetisation path for Muse is commissions: the company says it has no intention of profiting from compute subscriptions and plans to take a very small cut of transactions Muse facilitates (Zuckerberg, early September). On the night of September 20, however, Amazon blocked Muse from placing orders on its site, citing unauthorised access and agents that fail to identify themselves. For commissions to land, platform access has to be cleared first and user habits changed second. For now the subsidy's return shows up only in model capability, with neither subscriptions nor advertising in the recovery path.

IV. Free Tokens Displace Rentable Capacity, So Mark Them at List Price

Meta is preparing to sell hosted Muse Spark externally, which means every token given away to consumers already carries an external price. Bloomberg reported on July 1 that Meta Compute, led by infrastructure chief Santosh Janardhan among others, plans to rent out GPU capacity and to sell hosted-model calls in the manner of AWS Bedrock, Muse Spark included. On the July 29 second-quarter call, Meta disclosed that market compute prices sit well above its own procurement cost. Read together, the two data points establish that external demand is real. On that basis, every token the free tier burns is a token Meta does not sell at standard rates. Muse's cost should therefore be marked at standard pricing: roughly $13.7 of monthly opportunity cost per free weekly active user, about 7.5x the contributor-rate figure.

Accounting conventions make the subsidy look cheaper than it is. From fiscal 2025 Meta extended the depreciable life of certain servers and network equipment to 5.5 years. In practice, front-line GPUs have a first-cycle useful life of about three years, so the economic life of compute assets is shorter than the book life. On a three-year basis, annual economic depreciation on the same GPU is roughly 1.8x the book figure, which means any Muse cost computed off book depreciation systematically understates real consumption. Each GPU-hour Muse consumes also crowds out two alternative uses, ad ranking and external rental — and ad ranking is the internal use of compute with the clearest return at Meta.

An allowance that resets every week is effectively Meta's internal bid for its own compute. The quota can be reset weekly. When external rents rise, or when Meta Compute signs sizeable external contracts, the cost of maintaining a 100 million-token give-away rises with them. As an earlier report argued, Meta holds a long and a short position in compute simultaneously; Muse is the internal buyer on the long side, and the free allowance is its bid. We expect the allowance to be tightened, or shifted to weighted metering, once Meta Compute lands its first large external contracts. External prices already offer observable readings: the Ornn Data OCPI index settled H100 SXM rental at $2.97 per GPU-hour on August 31, and CME's H100 and B200 rental-index futures, built with Silicon Data, are scheduled to list on October 5 — giving the free tier's opportunity cost a tradable forward curve. Back to the opening question of which ledger takes the hit: part of the cost enters Meta's R&D expense and cash outflow, and part shows up as compute it never sold.

V. Investment Implications: R&D and Cash Flow Pay, Advertising Untouched

For Meta, the exposure runs through R&D expense and free cash flow, while advertising is shielded by the data separation. Second-quarter R&D expense was $21.66 billion, up 67% year over year, with free cash flow of $784 million. Applying the per-user costs above, at 10 million, 50 million and 100 million weekly actives the annualised free-tier subsidy comes to roughly $220 million, $1.10 billion and $2.20 billion at contributor rates, and roughly $1.64 billion, $8.22 billion and $16.45 billion at standard rates, the top end equal to about 19% of annualised second-quarter R&D. Apptopia estimates US mobile daily actives of about 642,000 on Muse's twelfth day, so the current subsidy remains comfortably within Meta's tolerance. The risk is that the subsidy scales linearly with weekly actives while subscription revenue does not. On valuation, multiplying subscriber counts by monthly fees to derive annual revenue ignores subscribers' own compute cost; Muse subscriptions should not be booked as SaaS recurring revenue. What tests this R&D spend is the next model generation's progress on agent benchmarks and the trajectory of Meta's third-party model spend. The New York Times reported on August 27 that, even after this summer's pullback, Meta still spends in the hundreds of millions of dollars a month on Anthropic model services; closing the agent gap in-house would create room to shrink that line.

For the compute rental market, how tightly Muse's quota is set determines how much supply Meta Compute can offer outside. The more inference capacity Muse absorbs, the less flows to external buyers through Meta Compute; a tighter allowance means more capacity reaching the rental market. Adjustments to Muse's quota can therefore be read as a leading signal for H100 rents and for the pricing environment facing independent compute clouds such as CoreWeave and Nebius. The effect concentrates in spot and newly signed contract prices, with existing long-term contracts largely insulated by locked-in pricing.

For consumer AI subscriptions, an ad-funded free agent drags down the reference price of the $20 tier. Muse's free tier is underwritten by the cash flow of Meta's advertising business, while OpenAI and Anthropic have no subsidy source of comparable scale on the consumer side. That puts pressure on both companies' individual subscription pricing. The exposure sits in consumer-facing price points; enterprise procurement turns on model capability and compliance terms and is far less directly affected by a free tier.

For China, the read-across lands on payments and merchant connectivity. Muse's commission path is hostage to third-party platforms' login and payment rails, and Amazon's block is the case in point. Tencent Holdings and Alibaba hold the agent entry point, the payment rail and the merchant network all at once; the link Muse lacks is precisely the one they own. Our July 7 report identified the integration of WeChat AI with WeChat Pay as the trigger for a winner-takes-all scenario, and Muse's experience at Amazon corroborates the value of that link from the other direction. For these two companies, the read-across is the transaction value embedded in payments and merchant systems, with only weak linkage to large-model subscription pricing.

VI. Three Tracking Lines: Quotas, Compute Rents, Model Capability

Three lines test the report's three conclusions. The quota line tests the “marketing unit” and “internal bid” claims: watch whether Meta publishes paid-tier allowances and a token definition, and whether changes to the free allowance coincide with Meta Compute's external deals. The rental line tests the opportunity-cost framework: watch the forward curve of the rental-index futures listing October 5, together with the price and tenor of Meta Compute's first deal. The capability line tests the data-purchase conclusion: watch progress against the 66.9% OSWorld 2.0 baseline from the next-generation model codenamed Watermelon and from later Muse Spark releases, plus the scale of Meta's third-party model spend. Two near-term nodes stand out: whether the September 23 Meta Connect keynote sets a timeline for Muse on AI glasses, and the R&D growth rate, free cash flow and Muse user metrics in the late-October third-quarter report.

Four developments would overturn or amend these conclusions. First, if Meta discloses that most paying users stay below the free cap and that subscription gross margin covers the free tier, Section II needs revising. Second, if Muse's privacy policy changes to let data into the ad system, the ad-acquisition characterisation becomes valid. Third, if the allowance is loosened after Meta Compute signs large contracts, the prediction that quotas tighten with rents fails, while the data-purchase conclusion is reinforced. Fourth, if agent benchmarks show no clear progress across two releases, or if the Confidential VM planned for this year becomes the default option, Meta's trainable data shrinks and the allowance most likely tightens.

VII. Risk Factors

Upside risk comes from falling costs and commissions materialising. Meta reports that Muse Spark 1.3 uses about 25% fewer tokens and about 20% fewer tool calls than 1.2; if internal inference cost keeps falling, the opportunity cost attached to the same allowance falls with it. Should utilisation stay low for an extended period, actual free-tier spend will come in below these estimates. And if commission revenue forms earlier than expected, Muse gains a recovery channel beyond subscriptions and R&D.

Downside risk comes from automation abuse, safety incidents and platform blocks. Round-the-clock automated tasks could concentrate quota consumption in a handful of accounts. Internal test records obtained by Reuters show Muse bypassing guardrails to retrieve photos from a tester's iCloud. If more platforms follow Amazon in restricting agent access, the commission path narrows further. Tightening the allowance is itself capable of triggering a reputational backlash.

The above content does not constitute any investment advice. Invest at your own risk.