Meta is back in the race for the most powerful AI models with Muse Wonder

Meta launches Muse Wonder and Wonder Full, two AI models designed for autonomous agents, with contextual memory, reasoning, and coding capabilities.

Meta is back in the race for the most powerful artificial intelligence models. The company introduced Muse Wonder and Muse Wonder Full, two new models designed for personal and professional AI agents. Developed by Meta Superintelligence Labs, they succeed the Muse Spark lineup and mark a new milestone in the group's ambition: creating assistants capable of understanding their users' needs and accomplishing complex tasks over time. Memory that stays with users

One of the main new features relies on Active Memory, a system designed to retain important information from one conversation to another and retrieve details that become relevant to a new request.

Meta also introduces Active Focus, a capability designed to keep the model's attention on a goal, even when a task requires multiple attempts. The agent can modify its approach, overcome certain difficulties, verify its work, and ask for confirmation before executing a major action.

Both of these mechanisms aim to move beyond the operation of assistants that merely respond to a sequence of isolated requests. Meta wants to develop agents capable of following a project, understanding its constraints, and continuing their work without having to start instructions over from scratch. Claimed performance against competing models

The new models were evaluated on professional tasks, particularly in law, finance, healthcare, and software development. Meta highlights their results on tests dedicated to long-term memory, programming, and multi-step task execution.

In its technical evaluations, the group positions Muse Wonder Full among the top-performing models on the market, notably against GPT-6 Astra and Claude Opus 5.5. Some external evaluations confirm top-tier performance, though results vary depending on the domain and measurement methods.

Meta illustrates these capabilities with several demonstrations: creating a 3D game from images, organizing a personalized running plan, arranging a wedding seating chart, or processing administrative procedures. Two models, different access levels

Muse Wonder is available starting today via the Meta Model API and Muse Code, while Muse Wonder Full, the most powerful version, is reserved for Muse Code subscribers. According to information published at launch, using Muse Wonder via the API is priced at $4 per million input tokens and $20 per million output tokens.

The model is also set to be integrated soon into Muse, Meta's personal assistant, with priority access for subscribers. Over the longer term, the company plans to rely on this family to develop specialized models intended for a broader audience. Toward more autonomous, but also more intrusive assistants?

With Muse Wonder, Meta continues its personal superintelligence strategy—an AI capable of accompanying users in their daily lives as well as at work. However, this evolution implies giving agents more context and, potentially, access to personal documents, messages, calendars, and applications.

Meta claims to have strengthened protections against malicious instructions and unauthorized disclosures of private data. Nevertheless, the effectiveness of these measures will need to be observed in real-world use, beyond the evaluations published by the company.

Above all, the release of Muse Wonder marks a scale shift for Meta. After rebuilding a large portion of its AI research infrastructure, the group now intends to compete with the most advanced models while preparing their integration into its own services.