Meta releases Muse Wonder AI models and adds new agent controls

In short: Meta has introduced two new AI models, Muse Wonder and Muse Wonder Full, and says the largest one competes with the world’s strongest models.

What happened

Meta announced Muse Wonder and Muse Wonder Full in a company blog post. Meta says Muse Wonder Full has “frontier” performance, meaning it sits near the top tier of today’s AI models.

The company says the models will be available through an API (a way for other apps to “plug in” to the model, like connecting an appliance to a power outlet). Meta also says Muse Wonder will be used inside its own products, including Muse Code, a tool that helps write software with AI, and features across Meta platforms, including its Muse personal agent.

Meta is positioning this as a result of a major AI reboot led by its Meta Superintelligence Labs. Mark Zuckerberg has invested billions and hired Alexandr Wang, previously CEO of Scale AI, to lead the effort. Some of the work has been rocky, with researchers leaving, and Meta has not shared details of how Muse Wonder was trained.

Meta claims Muse Wonder Full beats Anthropic’s Claude Fable 5.1 and sits just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index, a widely referenced score built from many tests. Meta also says it does better than OpenAI’s GPT-6 Astra on some specialized tests, including ones focused on legal documents and software building, though benchmarks do not always match real-world results.

Muse Wonder also includes “Active Memory” and “Active Focus,” which Meta says help AI agents stick to a goal and avoid getting confused (like keeping a long to do list without losing your place). Meta says it also trained the model to refuse requests for harmful instructions.

Why it matters

More capable AI models are increasingly being used as “agents,” meaning they can take multi step actions instead of just answering questions. If Meta’s new controls work, they could reduce costly errors and make these systems safer and more predictable for everyday users.

Source: Wired