← Pooja Verma
BenchmarkedIssue · 30 Sep 2026
Product launch · about a 6-minute read

Meta’s Muse: the most polished consumer agent yet, after a rough first three weeks

A free tier took Muse to #1 on both app stores. The model is mid-pack, and launch brought an Amazon block, a privacy dispute and a patched Mac flaw.

A plush Muse character pointing at an old desktop computer, illustrating the Muse Secure VM
Illustration from David Singleton’s (@dps) post on the Muse Secure VM, shared by @alexandr_wang.

The short version

  • A real errand-runner: Muse is an agent with its own cloud computer that emails, books, fills forms and shops, asking for approval first. vendor
  • Free, mostly: a free tier plus Power ($20/mo) and Maximum ($100/mo). Its two closest rivals have no free tier. press
  • A hit: #1 on the US App Store on Sep 18 and on Google Play on Sep 19, ahead of ChatGPT. independent
  • Mid-pack brain: Muse Spark 1.3 scores 45 on the Artificial Analysis index, #32 of 222 models. independent
  • A messy start: Amazon blocked it, a Mac flaw was hot-fixed, and an Inc. columnist says it synced his Messages after he declined; Meta says that access is opt-in. press

01What Meta launched

Muse launched in the US on Sep 8 on iOS, Android and muse.ai (TechCrunch also lists WhatsApp), with AI glasses “coming”. It takes long-running tasks, keeps working after you close the app, and asks for approval before it sends email or spends money. vendor

Watch the launch video on X (@Muse), where an order waits for approval of a $47.80 total, or Mark Zuckerberg’s live demo (via @RealNickMugalli).

02The security pitch

Meta’s answer to the trust question is architecture: each user gets the Muse Secure VM, a dedicated cloud computer, and a separate Sentinel agent must approve anything Muse sends to the internet. vendor

Meta's announcement: nothing Muse does reaches the internet unless the Sentinel approves it
The Sentinel, as described in Meta’s announcement. Source: Meta.

TechCrunch framed the trust question around Meta’s record, including a $5 billion FTC penalty in 2019. press

03The model underneath

Artificial Analysis model page: Muse Spark 1.3 (xhigh) scores 45, #32 of 222
Muse Spark 1.3 (xhigh) on the current Intelligence Index, v4.3.2. Source: Artificial Analysis, 30 Sep 2026.
MeasureMuse Spark 1.3 (xhigh)Kind
Intelligence Index45, #32 of 222independent
Rivals, same indexGPT-6.1 Sol 52 · Sonnet 5.5 56 · Opus 5.5 58 (all max)independent
Speed138.5 output tokens/sindependent
Tokens used140M (median 81M)independent
Cost per index task$1.37independent
API price (in / out per 1M)$1.25 / $4.25independent

The picture: decent, fast and very verbose, well behind the top tier. Muse’s edge is the product, not raw intelligence. Artificial Analysis tests the API model; the app may run a Muse-specific configuration.

04What people are doing with it

Meta’s Alexandr Wang started the #musemoneychallenge on Sep 12 (“can your muse make you $1000?”), and builders answered. vendor

Post by @KORGAONKAR, quoted by @alexandr_wang, about Muse retuning her Instagram feed
Muse retuned her Instagram toward companies offering gift cards for demos, then booked the calls. Source: @KORGAONKAR via @alexandr_wang.
Post by @mvanhorn, quoted by @MattPRD, about selling Warriors tickets with Muse
Muse answered buyers, tracked a spreadsheet and offered resale comps unprompted. Source: @mvanhorn via @MattPRD.

Adoption was fast: millions of downloads in about two weeks. Appfigures counts about 2.3 million; Sensor Tower and Apptopia figures cited in secondary reports are 3.4 million and 4.3 million. independent

05Amazon said no

TechCrunch quoting Amazon's error message about an unauthorized AI agent
The error Muse users saw on Amazon. Source: TechCrunch, via GeekWire.

By Sep 20, Muse users trying to shop on Amazon got an error calling it an “unauthorized AI agent”. Amazon confirmed it asked Meta to stop: Meta gave no advance notice, Muse doesn’t identify itself as an agent, and Amazon has concerns about how it handles customer credentials. press

There is a commercial angle too: agent shopping skips the browsing where Amazon sells ads. The two companies are also partners. press @Bencera summed up the standoff: “Either Amazon is building a Muse / Or Meta is building an Amazon.” community

06The Messages dispute

On Sep 19, Inc. columnist Jason Aten wrote that he explicitly declined access to his messages when installing Muse on an iPhone and a Mac mini. Muse later suggested that a chat with his podcast co-host would make a good column. Asked how it knew, Muse said it only saw notification banners relayed from the Mac. Aten says the Mac app had in fact synced his local Messages database, to row 187,462, while Full Disk Access showed as not enabled. press

Meta’s side: David Singleton, head of Meta Superintelligence Labs, said Messages access is opt-in and that Muse’s wrong explanation was “on us”, according to Decrypt. His reply to Aten implied the setting had been allowed. press

What’s unresolved: how the setting got enabled when Aten says he declined it. Meta hasn’t answered that, per Decrypt. This is one user’s account, in an opinion column; later AppleInsider and 9to5Mac pieces cover the same case.

07A patched Mac flaw, and humans on the phone

Gizmodo: an attacker with code already running on the Mac could redirect Muse's dictation
The precondition matters: the attacker already needed code on the Mac. Source: Gizmodo.

Security researcher Patrick Wardle disclosed a zero-day in the Muse Mac app on Sep 21. A hidden developer setting let malicious code already running on the Mac redirect Muse’s dictation to an attacker’s server, exposing the audio and an authentication token. Only the Mac app was affected. Meta removed the setting in a hot-fix shortly after midnight on Sep 22. Singleton called it “a local privilege escalation attack, not a remote exploit”. Wardle’s counterpoint: that precondition can be met by tricking a victim into pasting a malicious command. press

Separately, 404 Media reported on Sep 22 that, per internal messages, Meta “added a human agent layer” for Muse’s phone calls to businesses, and employee testers weren’t told a human placed the call until afterwards. Meta said this was internal testing ahead of public release. press

08Price, and how it stacks up

Meta: It's free for most of what people need, with subscription plans
Meta’s pricing line; per TechCrunch, Meta expects most users to stay on the free tier. Source: Meta.

Meta says Muse is “free for most of what people need”. vendor TechCrunch lists the paid tiers: Power at $20 a month and Maximum at $100. press Task-level pricing isn’t published, it’s US-only for now, and retail sites can block it.

Muse (Meta)dots (OpenAI)Grok Bot (SpaceXAI)
PriceFree tier, $20, $100From $100/moBundled at $120–300/mo
Free tier?YesNoNo
LaunchedSep 8Sep 29Aug 11 (beta)
ModelMuse Spark 1.3GPT-6 AstraGrok
StrengthDesign, free tier, WhatsAppSlack and Teams, 4,000+ appsBot teams, learns by demonstration

Dots pricing is OpenAI’s vendor; Grok Bot pricing is as VentureBeat reports it press.

09Handle with care

  • Download counts are third-party estimates and differ by firm; the safe statement is “millions, and #1 on both stores”.
  • Artificial Analysis’s Sep 2 launch article put Muse Spark 1.3 at 61–62 on an older index version. Those numbers aren’t comparable with today’s 45.
  • The Messages story is one user’s account. Meta says the access is opt-in; how it was enabled remains disputed.
  • The Mac zero-day required malicious code already on the Mac, and it has been patched.
  • Unconfirmed, so not reported here as fact: that the human-call program was “rolled back” or had “95–98% success” (neither is in 404 Media’s report), a $300K bug bounty, a Marketplace home-address claim, and a deleted post saying Muse made a user $2,500.

10Sources

Benchmarked · the AI news, fact-checked · Facts as of 30 Sep 2026 · labels show who reported each number: vendor, independent, press, community