Meta unveils Muse Wonder and Muse Wonder Full

What Meta announced

On 12 October 2026 Meta unveiled two new models, Muse Wonder and Muse Wonder Full. Its official announcement (opens in a new tab) says they deliver ‘state-of-the-art performance on personal and professional agentic tasks’. These are tasks where the AI acts as an agent, working through a request in several steps instead of answering a single question.

Meta gives examples such as insurance claims, lease reviews, financial questions and coding. It highlights two features: Active Memory, which carries relevant context across conversations, and Active Focus, which keeps working when a task gets hard, tries other approaches, recovers from errors and asks for clarification or confirmation.

Muse Wonder is available now through Meta’s Model API and Muse Code. Muse Wonder Full is for Muse Code subscribers only. According to the developer page (opens in a new tab), both have a context window of 1 million tokens, and Muse Wonder Full is included in the $50 and $200 a month subscription plans. Meta says Muse Wonder is coming soon to the Muse app, first for subscribers. The announcement gives no performance figures and says the full safety report will be available soon.

What it means for your business

For now this is news to follow, with no decision to take yet. Meta’s announcement includes no benchmark results, so there is nothing to judge it against for your kind of work. A model that scores well on coding tests will not necessarily write a better report for a local council or sort an accountancy firm’s email more accurately.

Three practical points. First, privacy: before you enter data about clients, staff or members of the public into any new tool, find out where it is processed, who can see it and whether it is used to train the model. Second, the EU AI Act and the GDPR: if the tool plays a part in decisions about people, such as recruitment or prioritising applications, you need stronger safeguards and human oversight. Third, vendor lock-in: you can trial a new model without changing your processes, and no single model should become something your work cannot do without.

Meta says the models avoid sharing private information without permission. That is the vendor’s statement, not a guarantee for your organisation. Your data policy remains your responsibility.

How to try it this week

• Pick a real task with little at stake, for example summarising a public document or drafting an internal announcement.
• Use only public or invented data. Do not paste names, email addresses, health data or confidential documents.
• Give the same task to the model you already use and to the new one, and compare the answers on the same criteria: accuracy, tone and the time it takes to check them.
• Check everything it returns before you use it. Note where it makes up facts or ignores instructions.
• Share the result with the team and decide together whether a wider trial is worth it, with agreed security rules.

IAescola designs tailored training for teams that want to adopt AI tools with good judgement, whichever vendor they choose.