Meta’s New AI Remembers More About You as Trust in Muse Wavers
Meta's new Muse Wonder models headline a feature called Active Memory, built to carry personal details across conversations.
Meta introduced Muse Wonder and Muse Wonder Full on Monday as its most capable models yet and the first from Meta Superintelligence Labs built to compete with OpenAI and Anthropic’s frontier systems.
The launch caps what Meta called a 14-month overhaul of its AI effort, building on April’s Muse Spark model.
Muse Wonder is available via the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, matching Anthropic’s Claude Opus 5.5 pricing.
Muse Wonder Full, the more capable version, skips the API entirely and is available only to Muse Code subscribers, requiring a subscription rather than pay-per-use access.
Memory and Follow-Through, Not Just Raw Intelligence
Meta built the launch around two specific capabilities rather than generic intelligence claims.
Active Memory carries relevant details across conversations, so a user doesn’t have to re-explain context Muse has already learned.
Active Focus is meant to solve a different problem: models that give up or wander off course on long, difficult tasks, by having Muse try alternate approaches, recover from errors, and explicitly check in with the user before following through on anything consequential.
Meta is also expanding into expert domains, law, finance, and health, positioning Wonder to execute complex tasks like interpreting leases or processing insurance claims rather than an everyday chatbot answering questions.
The Benchmarks Don’t Quite Match the Marketing
Meta’s own announcement leans on language like “state-of-the-art” and “frontier,” but independent testing, as Mashable notes, complicates that framing.
Artificial Analysis, which runs controlled comparisons across major AI labs, found Muse Wonder Full ties Anthropic’s mid-tier Claude Sonnet 5.5, not its flagship Opus 5.5, which still holds the top overall ranking ahead of Wonder Full, GPT-Astra, and Gemini 4.
That’s a very different claim than “our most capable model yet” implies to a reader skimming the headline: Meta’s newest, most expensive model matches a competitor’s second-tier offering, not its best one.
Some of Meta’s own benchmark comparisons were also run without the kind of controls that let outside observers verify them directly.
A Memory Upgrade Lands in the Middle of a Trust Problem
A few Weeks before this launch, Muse accessed an Inc. columnist’s private Mac Messages without explicit permission, sparking a debate with Meta and prompting Apple to tighten macOS privacy controls.
Active Memory enables an AI agent to retain personal context across conversations, exactly the capability that raises the stakes if an agent’s permission handling isn’t fully trustworthy yet.
Meta’s announcement acknowledges this tension, stating models were trained to avoid sharing private information without permission, wording that feels less like a feature highlight and more like a response to recent criticism.
Shipping deeper memory and autonomous follow-through is a logical bet if trust is restored, but it remains unclear whether Meta has solved that problem or merely restated its goal.
Source: Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents
