Analysis
Meta launched Muse this week, a personal AI agent meant to act on a user's behalf across email, calendar, shopping and the smart home rather than just answer questions, The Verge reported. Muse runs on Muse Spark, an early model from Meta Superintelligence Labs, and ships in three tiers -- free, $20 a month and $100 a month -- available for now only to US users 18 and older, through a standalone app and inside WhatsApp.
Meta says Muse already reaches into a working list of outside services:
- Gmail -- drafting and sending email
- Spotify -- managing playlists and subscriptions
- Ticketmaster -- buying event tickets
- OpenTable -- booking restaurant reservations
- Shopify -- placing orders on connected stores
The launch is Meta's answer to a rough year in AI. Llama 4's Scout and Maverick models landed to a flat developer reception in April, and later reporting showed Meta's own team had cherry-picked which model version got submitted to different benchmarks -- including submitting a different version to the LM Arena leaderboard than the one developers actually received as open weights. Mark Zuckerberg responded by paying roughly $14.3 billion for a 49% stake in Scale AI to bring in its CEO, Alexandr Wang, as Meta's chief AI officer, then built Meta Superintelligence Labs around him with signing bonuses reported as high as $1 billion for top researchers and an equity package for Wang alone estimated at $100 million to $150 million.
Muse puts Meta squarely against agent products every other major lab has already shipped: OpenAI's ChatGPT agent mode books and buys with minimal supervision, Google's Gemini ships Project Mariner as a browser agent, and Anthropic markets Claude's computer-use features for multi-step business tasks. None of those companies has Meta's distribution -- Muse rides directly into WhatsApp, which counts more than 2 billion users -- but none of them spent the past year getting caught cherry-picking its own benchmarks, either.