Meta launches Muse Wonder and Muse Wonder Full: a frontier model family built for personal and professional agents

Start with what the official post actually says. Meta's newsroom post is titled 'Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents' and dated October 12, 2026. Four things are central. First, the two models are Meta's most capable to date, aimed at state-of-the-art performance on personal and professional agentic tasks, with examples running from navigating an insurance claim and making sense of a lease to building a working app. Second, availability: Muse Wonder is available from launch to developers and builders through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers; the family will soon power experiences in Muse, rolling out first to subscribers. Third, the technical lineage: Meta says it began an overhaul of its AI efforts fourteen months ago and that the Muse Wonder series builds on groundwork laid with Muse Spark. Fourth, safety: Meta says the models were trained to avoid sharing private information without permission, to turn down requests to assist in actually harmful acts, and to help improve security and strengthen defences against cyber attacks. Meta adds that it trains, evaluates and releases its most capable models in line with its Meta Superintelligence Scaling Framework, and that a full report with additional safety evaluations will be available soon.

Two new capabilities deserve their own section, because they explain why this family is positioned as agentic. The first is Active Memory, defined officially as the ability to carry what matters to the user across conversations and bring back the details relevant to the task at hand. Memory, in other words, is no longer confined to a single session. The second is Active Focus, defined as the ability to keep working toward a goal when a task gets hard: trying other approaches, recovering from errors and following through, asking for clarification when needed and checking in before following through on anything important. Meta's own framing is that real work unfolds over time, so a model has to retain context across conversations, and that real work hits dead ends, so a model has to be able to switch approaches instead of stalling. For ordinary users these definitions decide whether an agent can be trusted with long tasks; for developers they also imply new failure modes — a memory that carries private information it should not, or an Active Focus that works around limits that were there for a reason.

Pricing, tiers and outside evaluation are the more practical parts of this launch. CNET reports that Muse Wonder is Meta's premium-tier model, available first to coders who pay for access to the Meta Model API through a Power Plan at $20 per month or a Maximum Plan at $100 per month, while Muse Wonder Full has no public API and is available only to developers who sign up for the paid Muse Code agent, bundled into a subscription priced from $50 to $200 per month. CNET notes that Meta's announcement says Muse Wonder will soon power Muse experiences with more flexible per-million-token pricing. On independent evaluation, Mashable reports that the benchmarking organisation Artificial Analysis placed Muse Wonder Full second only to Claude Opus 5.5 Max and tied with Claude Sonnet 5.5 Max on its Intelligence Index, ahead of OpenAI's GPT-Astra, Anthropic's Fable 5 and the newly released Gemini 4. Artificial Analysis said both models show a material jump in overall Intelligence Index score, with Muse Wonder Full placing in the top three on six of ten evaluations and being particularly strong in agentic knowledge work and scientific reasoning. Mashable also supplied the timeline: Muse Wonder lands just six weeks after Muse Spark 1.3, following heavy Meta investment in AI recruiting and infrastructure, including the $14.3 billion acquisition of Scale AI.

Finally, safety and the surrounding argument, which should not be buried under product news. Meta's post says a full report with additional safety evaluations will be available soon, which means the model is open to developers while the complete safety evaluations are not yet public. CNET names the tension directly: the industry is in a period of heightened safety attention, researchers and experts are pushing for external evaluation and regulation, and Meta CEO Mark Zuckerberg has argued publicly that labs can and should self-police, telling the Financial Times this week that he is much closer to believing AI can be engineered with normal guardrails and built responsibly than to the anthropomorphic vision of AI doing its own thing. CNET's reporter also notes that Active Memory means storing potentially sensitive data, which should put privacy-conscious users on edge, and flags Meta's claim that developers and security teams can use models like Muse Wonder to find flaws in their own code as a claim worth questioning. One more piece of context: Anthropic previously revoked access to Mythos 5 and Fable 5 after the U.S. government imposed sweeping restrictions on who could use them, which shows frontier model availability is no longer purely a commercial matter. These are outside readings of the same announcement; the official position in the post is that the models are released in line with Meta Superintelligence Scaling Framework standards.