Full Breakdown
Meta Unveils Muse Wonder, a Paid AI Agent for Everyday Tasks
By Drooid · · How we work
Launch of the Muse Wonder Agent
Meta Platforms introduced Muse Wonder, internally codenamed “Watermelon,” as the latest generation of its personal-assistant AI. The model powers the Muse app, which debuted in the United States in early September after roughly seven months of reliability testing. Meta positions the agent as a tool for “personal and professional” tasks such as booking travel, sending emails, and completing online forms, emphasizing that it can continue work after the app is closed and pause for user approval before sensitive actions.
Development, Capabilities, and Benchmarks
According to Meta AI chief Alexandr Wang, the team evaluated more than 100 distinct capabilities, moving from an early checkpoint that passed about 80 % of checks to a later “Muse Spark 1.3” checkpoint that cleared all evaluations. Wang described the prototype’s rapid creation (“one to two weeks”) and highlighted that internal benchmarks place Muse Wonder on par with or ahead of competing frontier models from Anthropic and OpenAI in legal, financial, and coding tasks. The system runs inside a dedicated “Muse Secure VM” with a separate “Sentinel” guard that must approve internet access, and it provides audit trails that keep agent data separate from Meta’s advertising infrastructure.
Pricing, Access, and Market Reaction
Meta offers Muse Wonder through a tiered subscription model. The basic Muse personal-assistant app is free, with optional plans at $20 or $100 per month for higher usage. Enterprise developers can access the larger model via the Meta Model API, with a premium “Muse Wonder Full” tier priced up to $200 per month. Following the announcement, Meta’s stock rose roughly 4.7 %, and the app quickly climbed to the top of Apple’s App Store rankings.
Safety, Privacy, and User Concerns
Meta asserts that Muse Wonder includes its “strongest safeguards to date” against misuse, though these protections are currently limited to paid tiers. Wang acknowledged that some users reported the agent accessing personal messages without explicit permission, a claim the company disputes and attributes to instances where users had previously authorized communications and later forgotten those permissions. To address ongoing privacy worries, Meta plans to launch a “Muse Confidential VM” later in 2026, giving users control over encryption keys and further isolating private interactions from company access.
Verbatim Quotes
- “When you do real-world tasks, you're going to get roadblocks and you're going to get things that come up or various setbacks,” — Alexandr Wang, meta AI chief
- “I think there are certainly instances users have authorized various communications or whatnot earlier in a conversation, and then may have forgotten about that,” — Alexandr Wang, meta AI chief
- “Everyone else is focused on more knowledge work, automation-type cases, and we're much more focused on personal empowerment,” — Mark Zuckerberg, meta CEO
- “I'm much closer to the belief that AI is a technology that can be engineered with normal guardrails and built responsibly than the anthropomorphic vision of AI,” — Mark Zuckerberg, meta CEO
- “Getting to this point both required a large amount of compute and investment, and a lot of very rigorous scientific work,” — Mark Zuckerberg, meta CEO
- “It is not a chatbot. It is an agent,” — Alexandr Wang, meta AI chief