Meta hands Muse users 100 million free tokens a week

Meta's own copy for Muse asks users to "Take the busywork off your plate" ^[3]. What Gizmodo's writer handed it instead was an author photo redrawn line by line in Python, because the agent's Linux terminal had no MS Paint equivalent, a MIDI tune called "Portrait Theme", a side-scrolling game, an FPS variant with NPC versions of the writer, and a retro website called "Garden Signals" to hold all of it ^[4]. He was clear about why: he had no busywork he trusted an agent to do ^[5]. When he pushed for volume, asking for 50 genre variations of the tune, what came back was mostly the same melody with the synthesized instrument swapped, the "Dubstep" version barely separable from the "Sea Shanty" one ^[6].

That gap between what the page promises and what the session produced is the useful part. Teams tell themselves a free agent tier goes on ticket triage and release notes. The one documented user built things nobody had to check, because checking was never the point.

The ceiling is worth doing the arithmetic on. A hundred million tokens a week is about 14.3 million tokens a day ^[13], roughly 5.2 billion a year if the allowance holds ^[14]. One person typing prompts at an agent on a virtual machine ^[15] will not reach that, which is exactly what Gizmodo found while trying to ^[2]. Where a long-running agent loop sits against the same cap is not in the account, so nobody should treat this as a capacity test for production work.

The disposal reading has support, though not proof. Gizmodo reports that in July Meta was looking to go into business as a cloud provider for other companies, and that critics at the time called it a sign the company had overbuilt its data centres ^[7]. A projection in April put Meta's capital expenditure at $145 billion this year ^[8], and free cash flow fell 91 percent from the first quarter to the second ^[9]. Zuckerberg's framing on last month's earnings call was demand, not surplus: usage is ramping, so infrastructure investment continues ^[10]. Both readings fit a very wide free tier. What tilts it slightly is that Meta already spent a few months telling its own staff to consume as much AI compute as they could ^[11], which is a thing you do with capacity you have already paid for.

The part of this that does not survive contact with the evidence is price. No per-token rate for Muse or anything else appears in the account ^[16], so the giveaway cannot be used to argue that agentic inference should now cost less. You cannot plan a budget around zero. The number that decides whether a Muse workflow lives is the cap after the promotion ends, and the current one exists as a Threads post from Mark Zuckerberg ^[1].

There is also a question of who shows up. Alex Kantrowitz of Big Technology argued on CNBC in March that Meta's billion monthly AI users were bumping into AI inside existing apps rather than seeking it out, and that its intentional users were nothing close to ChatGPT's 900 million weekly ^[12]. Muse requires intent. Somebody has to decide to hand over real work.

Two questions decide whether that free tier is worth using on your own team: whether you can verify the output faster than you could have produced it, and whether it breaks anything downstream if it is wrong. Fast to verify and reversible is where a free agent tier earns its place: draft sites, throwaway tooling, one-off scripts. Slow to verify and irreversible is where the free tokens are the most expensive thing in the building, because the review time is yours and the allowance is not. The rewrite is the real cost here, not the tokens.