Meta has introduced Muse Wonder and Muse Wonder Full, two AI models designed to power agents that can work across longer, more complex personal and professional tasks. The company says the models improve on earlier systems by retaining relevant context, persisting through setbacks, and taking a more deliberate approach to tasks such as coding, planning, research, and document analysis.
Muse Wonder is available through Meta’s Model API and its Muse Code product. It will also begin rolling out in the Muse personal agent app for paying subscribers. Muse Wonder Full, the more capable version, is initially limited to Muse Code subscribers.
The launch is part of Meta’s effort to turn its AI work into subscription and developer revenue beyond its advertising business. Jonathan Vanian reports for CNBC that access to Muse Wonder in the consumer app requires a $20 or $100 monthly subscription. Muse Wonder Full is tied to Muse Code plans priced at $50 or $200 per month. Meta AI chief Alexandr Wang tells CNBC that API access will cost more than standard subscriptions, though Meta intends to price it below comparable leading models from OpenAI and Anthropic.
Memory and sustained work
Meta frames the new models around two concepts: Active Memory and Active Focus. Active Memory refers to a system’s ability to retain information that matters to a user across conversations, then retrieve it when it becomes relevant. In practice, that could include preferences, prior decisions, calendar information, or details from documents that the user has permitted the agent to access.
Active Focus concerns an agent’s behavior when a task does not work as expected. Rather than stopping at the first obstacle, Meta says Muse Wonder can try alternative methods, incorporate new information, and ask for clarification before important actions. Wang describes this as a response to common real-world problems, such as a booking system with no available reservation times.
Meta says the models perform strongly on internal and industry benchmarks covering legal, financial, health, memory, coding, and long-running software tasks. The company also argues that Muse Wonder Full transfers capabilities more effectively to unfamiliar tasks than its earlier Muse Spark 1.3 model. Those claims are based on Meta’s own evaluation framework and should be viewed alongside independent testing as it becomes available.
For content and knowledge workers, the key promise is less about answering single prompts and more about handling multi-step work. Meta presents examples including preparing a seating chart with detailed constraints, creating a running plan based on personal circumstances, and working across software projects over extended periods.
Privacy and safety questions
More capable agents also raise higher privacy stakes. Meta says Muse Wonder includes its strongest safeguards so far against cyber and CBRN misuse, referring to chemical, biological, radiological, and nuclear risks. The company also says it trained the model to handle private information carefully and resist adversarial instructions.
That message comes as Meta faces scrutiny over how its Muse agent accesses user data. Some users have alleged on social media that Muse accessed personal messages without permission. Meta disputes those claims and says some users may have previously authorized access during earlier conversations. Wang says the company is monitoring agent-related incidents and improving guidance for users.
Sources
- Meta introduces Muse Wonder and Muse Wonder Full, saying they deliver state-of-the-art performance on personal and professional agentic tasks – Meta AI
- Meta says Muse Wonder is available to Muse app subscribers, via Meta Model API, and Muse Code; Wonder Full is available to Muse Code subscribers – CNBC
- Meta says it plans to power its Muse AI agent with Wonder, starting with subscribers, and will develop specialized models it “can bring to everyone using Muse” – The Verge
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
