In early September this year, Meta (META.O) unveiled an AI tool called Muse in the United States and Canada. Within two weeks, it rocketed to the top of the U.S. App Store’s free app chart, dethroning ChatGPT, which had dominated the rankings for a considerable time. This achievement was no small feat. According to Sensor Tower data, Muse racked up over 2.5 million downloads in its first 13 days across the two major markets. During the same period, ChatGPT recorded 3.1 million global downloads—though it’s worth noting that ChatGPT launched globally on iOS, while Muse is currently only available on iOS and Android in the U.S. and Canada, making a direct comparison somewhat imprecise.

Nonetheless, Muse’s downloads far outstripped Claude’s 400,000 and Grok’s 200,000. More importantly, its growth rate was remarkable: Muse averaged a 55% daily increase in downloads in its first two weeks, compared to ChatGPT’s 24% during the same period. By late September, cumulative downloads exceeded 3.4 million, with daily active users numbering approximately 700,000, according to Similarweb. The capital markets responded with even more direct enthusiasm.

By late September, Meta’s stock price had surged by over 25% since Muse’s release. On September 21, it rose over 13% intraday, closing up 11.43%—its largest single-day gain in more than a year—adding roughly $190 billion to its market cap in a single day and bringing its total valuation to nearly $2 trillion. In a late-September interview, Mark Zuckerberg described the early reception of Muse as a “home run right out of the gate.” But focusing solely on download numbers would overlook the real secret behind Muse’s explosive popularity.

01. Distribution Dominance with 3.6 Billion Daily Actives—A Moat Others Can’t Replicate

An often-overlooked fact is that Muse is not built on the industry’s most cutting-edge model. According to third-party evaluation agency Artificial Analysis, Muse Spark 1.3 (xhigh), which powers Muse, scored 61 on its intelligence index, tying with GPT-5.6 Sol (max) but still trailing Anthropic’s Claude Fable 5.1 (66). By traditional standards, this shouldn’t be a hit. But it is, and for good reason.

Muse and ChatGPT are fundamentally different. ChatGPT is a “conversationalist”—you ask, it answers, with core interactions revolving around question-and-answer rounds. Muse is an “executor”—you tell it a task, and it fires up its cloud-based computer, navigates browsers, fills out forms, completes bookings, and even places orders and payments on your behalf. It continues working in the background even after you close the app.

At Meta Connect 2026 on September 23, Meta’s Chief AI Officer, Alexandr Wang, and Zuckerberg set the tone for Muse. Wang noted that a year ago, “running a personal AI agent on your own computer” was the domain of tech enthusiasts, but “today, we’re bringing that magic to everyone through Muse.”

Zuckerberg set an even grander vision: Muse will evolve into a “personal superintelligence used by billions”—capable of remembering your preferences, personalizing your actions, and orchestrating cross-app workflows. In other words, Muse’s popularity stems first from turning “AI doing things for you” from a concept into a product ordinary people can actually use. No complex prompts, no need to learn “how to talk to AI”—you just tell it what you want. The barrier to entry is nearly nonexistent. If product positioning is Muse’s internal strength, Meta’s distribution power is its external supercharger.

According to third-party agency Apptopia, over 95% of Muse users are also Facebook users, and 63% use Instagram. Meta’s global product family had 3.6 billion daily active users (DAP) by the end of Q2. This means Muse only needs cross-promotion within Meta’s ecosystem to reach this massive existing user base—though its services are currently limited to the U.S. and Canada.

China Merchants Securities International concluded in a research note that Muse is tapping into a personal AI agent market shaped by multiple entry points, and Meta’s app family with over 3 billion daily actives remains the largest ready-made distribution channel in this market—large platforms can deliver resident AI agents to billions of users at nearly zero acquisition cost. This is a structural advantage independent AI apps cannot replicate. While OpenAI needs to spend billions on traffic acquisition, Meta only needs to add an entry point on Facebook or Instagram.

02. An ‘Always-On’ Virtual Computer—The Source of Security

What truly makes people willing to entrust Muse with their passwords lies in its architectural layer. Meta assigns each user an independent cloud-based virtual machine with 2 virtual CPUs, 8GB of memory, and 100GB of solid-state storage, complete with its own browser and file system. This virtual machine runs all authorized account connections—email, calendars, payment info.

It also features a security mechanism called Sentinel, which requires both system approval and user confirmation before the agent executes sensitive operations. This addresses the biggest psychological barrier for consumer-grade AI agents: trust. To have an AI book hotels, modify orders, or even spend money on your behalf, users need more than “it’s smart”—they need assurance that “even if something goes wrong, it won’t spiral out of control.”

The independent virtual machine architecture provides this boundary—it acts like an isolated assistant with its own workspace, never intruding on your device. As Geek Park wrote in a product retrospective, Muse’s design team believes agents need to be proactive without being annoying, capable of taking action without losing control, and knowledgeable about everything you do while remaining trustworthy enough to hand over your data. Balancing these three contradictions is where Muse’s true product strength lies.

03. The Pain of Computing Power: The Cost of Explosive Growth is Emerging

Muse isn’t without challenges. Before daily actives even hit a million, its servers were already flashing red alerts. Meta’s Chief AI Officer, Alexandr Wang, revealed on September 9 that Muse’s actual usage far exceeded internal projections, with user consumption roughly 10 times that of the beta testing group. Third-party status monitoring platform SaaSHub has labeled Muse as “degraded,” with users continuously reporting issues like “unable to search.” During a stress test, Muse attempted to create 120 sub-agents simultaneously but succeeded with only 33.

The root problem lies in the independent virtual machines—each additional user means another virtual machine to run. Industry estimates suggest that if user scale reaches 100 million without resource sharing among users, Meta would need at least 200 million CPU cores, equivalent to roughly 1.58 million 126-core AMD EPYC processors, plus 800PB of memory and 10,000PB of solid-state storage.

And Muse is currently only available in the U.S. and Canada, with WhatsApp and Instagram registration entry points still offline. Once fully rolled out, the computing power gap will widen further. This is a sweet problem, but a problem nonetheless. Whether Meta can transform Muse’s growth momentum from a “hit product” into a “sustainable business” won’t be found in download numbers—but in the racks of its data centers.

04. The Second Half of the AI Race: The Winner May No Longer Be the Smartest Model

Muse’s explosive popularity reveals an industry trend: When most models reach “good enough” capability levels, the ability to package technology into usable products, push them to mainstream users, and continuously iterate based on feedback is becoming just as important as model parameters themselves. Meta wasn’t the first to propose the “AI agent” concept. OpenClaw gained traction among developers earlier this year, but it required technical background to deploy, and misconfigurations could lead to agents acting out of bounds—in February, an AI safety lead at Meta’s superintelligence lab had over 200 emails deleted in bulk after an agent lost safety instructions during context compression.

What Muse does is repackage this capability to make it usable for ordinary people who have never interacted with agents before. Zuckerberg has also clearly outlined Muse’s business model: The free tier offers approximately 100 million tokens per week plus one cloud virtual machine; heavy users can upgrade to $20/month (500 million tokens/week) and $100/month (3 billion tokens/week) subscription tiers.

But he places greater emphasis on another path—future small commissions from transactions facilitated by Muse, letting AI truly “spend money on behalf of users” and derive value from consumer spending flows. To be clear, transaction commissions are still in the planning stage and not yet implemented; subscription revenue remains the primary revenue stream for now. Whether this path works depends on whether Muse can truly become an indispensable assistant in the daily lives of billions. Based on current data, it has at least taken the first step. And the next phase of competition in this industry may no longer be about whose model scores highest on benchmarks, but who can truly make AI “get things done for people.”

- End -