Freemium With a Compute Bill: What Meta Muse’s 100 Million Free Tokens Teach

Meta gives most Muse users the agent for free and says it is very compute intensive to run. Here is what that choice shows founders about freemium when every free user carries a real inference bill, and why an ad giant can make a bet a startup usually cannot.

When Meta launched its Muse agent on Sept. 8, 2026, it made most of the product free, even though each agent runs on its own cloud computer. On its Oct. 21 earnings call, Zuckerberg told investors that Muse “is very compute intensive to run.” For founders, that pairing is the cleanest live case study of freemium AI compute costs: the free tier is not a marketing line item, it is an infrastructure bill that grows with every sign-up.

Where it stands: Meta has not published what Muse costs to serve. Asked on the Oct. 21 call for per-user cost metrics, CFO Susan Li answered that “we do not have sort of specific numbers to share around the cost structure of Muse today, but, of course, it is, as Mark said, pretty compute expensive.”

What makes freemium different when the product is an AI agent?

In classic software, a free user costs a little storage and bandwidth. In an AI agent, a free user triggers inference, the model computation that runs every time the AI reads a request and writes an answer, and that work is paid for in GPU time.

Muse sits at the expensive end of that spectrum. Meta’s launch post says the agent “runs on its own dedicated computer in the cloud, contained so no one else’s agent can reach it,” and that “for tasks that take more time, Muse keeps working after people close the app.” A chatbot answers and stops. An agent like this can keep a machine busy while the user is asleep.

PYMNTS made the same point in September, writing that the agent’s “compute bill per session is higher than a chatbot’s.” That is PYMNTS’s characterization, not a Meta figure.

How big is the Muse free tier?

Large, by any consumer-software standard. Gizmodo reported on launch day, citing a Threads post by Zuckerberg, that Muse “will currently let you use 100 million tokens per week for free.” A token is a small chunk of text, roughly a word fragment, that a model reads or writes; vendors meter AI usage in tokens.

Meta’s own newsroom post does not give the number. It says only that Muse is “free for most of what people need, with subscription plans for people who want to do more.” TechCrunch reported the two paid plans at launch: “Power at $20/month and Maximum at $100/month.”

MarketBeat’s Leo Miller spelled out the trade-off from the investor side: “every token Muse consumes without payment is simply a cost to Meta, which could damage near-term profitability.” That is an analyst’s view, but the logic applies to any founder pricing a free AI tier.

Design choice What Meta did with Muse Reported by
Free allowance 100 million tokens a week Gizmodo, citing Zuckerberg’s Threads post
Paid upgrades Power $20, Maximum $100 per month TechCrunch
Card at signup Required to get started TechCrunch
Usage visibility In-app meter of allowance used Gizmodo
Cost per user Not disclosed Susan Li, Oct. 21 call

Why can Meta subsidize a free tier that a startup cannot?

Because the money comes from somewhere else. Li told analysts that “Q3 Family of Apps ad revenue was $63.1 billion, up 26% on both the reported and constant currency basis.” Zuckerberg framed Muse’s runway in exactly those terms: “In the near term, the strength of our core business enables us to scale the Muse network before then.”

A writer for TIKR, published on Yahoo Finance, put it more bluntly as commentary: “It comes down to economics: Meta can subsidize the demand because of its enormous ad business, while OpenAI and Anthropic cannot.” Treat that as opinion. It is still the core lesson for a seed-stage company. A free AI tier is financed either by another profitable business line or by investors’ cash, and only one of those refills itself every quarter.

Meta also has a stated plan for who eventually pays. On the call Zuckerberg said: “For most people, the product should be free, and we will build commerce tools so we get paid a fee based on the transactions that we help facilitate. We expect this will be paid for by businesses.” The free user is meant to be monetized through merchants, not through a paywall. Our sibling piece on growth before monetization looks at that sequencing in detail.

What guardrails did Meta build into the free tier?

Three design details stand out, and each one can be copied by a small team.

A payment card up front. TechCrunch reported that Muse “will be free to use, with subscription plans kicking in as usage increases, which is why Muse requires a payment card to get started.” Asking for a card filters out throwaway accounts and makes the upgrade path frictionless when a user hits the ceiling.

A visible meter. Gizmodo’s reporter tested the free tier and quoted the agent telling him: “You’ve used about 6% of your free weekly allowance, and it resets on September 15.” Showing the meter turns a hidden cost into a budget the user can see.

A weekly reset. A weekly allowance caps how much compute any single free account can burn before the counter starts over, which a monthly or lifetime allowance does less tightly.

Meta’s AI chief Alexandr Wang also described the paid plans as a cost tool. “Those subscription tiers help us cover the compute costs,” he told Axios, as quoted by PYMNTS.

How did rivals contain freemium AI compute costs from heavy users?

Several AI companies pulled back from generous flat pricing once power users showed up, and those episodes are the warning label for any generous free tier.

In July 2025, TechCrunch reported that Anthropic was adding weekly rate limits for subscribers running Claude Code “continuously in the background, 24/7.” Anthropic spokesperson Amie Rotherham told TechCrunch the limit would affect “less than 5% of subscribers, based on their current usage patterns.” A small slice of users was enough to force a policy change.

Cursor, the AI code editor, explained its own switch in a company blog post: “the hardest requests cost an order of magnitude more than simple ones. API-based pricing is the best way to reflect that.” Usage-based pricing is a deep topic of its own; we cover Cursor’s switch and Muse’s token tiers in our usage-based pricing guide.

The common thread is that averages mislead. A free or flat tier priced for the typical user can still lose money if a few accounts behave very differently from the median.

What should founders copy, and what should they skip?

Copy the mechanics. A card at signup, a visible meter, a short reset window and a clear upgrade path cost little to build and protect your margin. Our explainer on margin after inference costs walks through how to measure what each account actually consumes.

Be careful about copying the size of the allowance. Meta’s bet rests on an expectation it stated on Oct. 21: “we expect to improve its cost efficiency by more than 10x in the next year, which will enable it to be profitable once the revenue lines are at scale.” That is a company forecast, not a result, and it is backed by an advertising business few startups will ever have.

A smaller company can test the same freemium model with a much tighter free allowance, raise it as its own serving costs fall, and keep the free tier as a funnel rather than a destination. For background on the model itself, see what the freemium pricing model is.

Meta’s own description of the paid tiers’ job came from Wang in September: they exist to “help us cover the compute costs.”

Frequently Asked Questions

Does freemium work for AI products?

It can, but each free user costs real compute. Meta offers Muse free for most people and sells $20 and $100 monthly plans, while saying the agent is very compute intensive to run. Meta has not disclosed what Muse costs to serve.

How many free tokens does Meta Muse give away?

Gizmodo reported on Sept. 8, 2026, citing a Threads post by Mark Zuckerberg, that Muse lets free users consume 100 million tokens per week. Meta’s own newsroom post does not list the allowance.

Why can Meta afford a generous free tier when startups cannot?

Meta reported $63.1 billion of Family of Apps ad revenue for the third quarter of 2026, and Zuckerberg said the core business lets Meta scale Muse before revenue arrives. A TIKR writer argued that OpenAI and Anthropic cannot subsidize demand the same way. That is the writer’s opinion.

How do AI companies limit heavy free or flat-rate users?

Common tools are hard caps, usage meters and weekly limits. Anthropic added weekly limits in July 2025 that it said would affect fewer than 5% of subscribers, and Cursor moved its Pro plan to a pool of included usage.

Sources

This article is general information for founders and is not accounting, legal or investment advice. Figures are as reported by the sources cited on the dates shown; analyst and third-party estimates are not company disclosures. Found Founders is not affiliated with Meta.