TL;DR

Meta released Muse Wonder and Muse Wonder Full on Monday 12 October, describing them as the strongest models it has built and the product of an AI overhaul that began fourteen months ago. Meta stock gained around 4% during afternoon trading, Reuters reports. The launch post gives no benchmark figures, and a report with further safety evaluations is promised rather than published, so the claims cannot yet be checked against OpenAI’s and Anthropic’s latest releases.

What Meta is claiming

The new models build on Muse Spark, the first release of the rebuilt effort, and are pitched squarely at agent work: reading an insurance claim, parsing a lease or sorting out someone’s finances, plus coding. Meta says they reach expert level in law, finance and health, and frontier level in code, without saying how that was measured.

Two features carry the marketing. Active Memory is meant to hold what matters to a user across conversations and surface it when relevant. Active Focus is Meta’s name for persistence: trying other approaches when a task stalls, recovering from errors, asking for clarification, and checking in before anything consequential.

Muse Wonder is available now via Meta’s Model API and inside Muse Code, the company’s coding tool, and will roll into the Muse personal agent, subscribers first. According to Reuters, Muse Wonder Full, the larger of the pair, goes to Muse Code subscribers.

Safety, briefly

Because an agent of this kind reads messages, calendars and documents, Meta says the models were trained not to share private information without permission, to refuse help with harmful acts, and to support defensive security work, including letting developers probe their own code for flaws. It says the models were trained, evaluated and released under its Superintelligence Scaling Framework, with the fuller evaluation report to follow.

Looking forward

Reuters frames the launch as the payoff from Mark Zuckerberg’s year of heavy spending and aggressive researcher hiring, and notes that the Muse agent’s rollout last month gives Meta a subscription route to recoup it, privacy concerns notwithstanding. It is the third frontier release in a week, after OpenAI pushed GPT-6 to every ChatGPT user and Anthropic cut the running cost of Claude Haiku 5.5. For UK businesses, the practical question is not whether Meta has caught up but what the delayed safety report says about an agent built to work inside inboxes and calendars. In our view, a launch post that promises evaluations later is a reason to wait, not to switch.