Meta pushes Muse Wonder from apps into agent infrastructure

What Meta says the models can do

Meta’s announcement frames Muse Wonder and Muse Wonder Full as models for “real world” work: navigating an insurance claim, understanding a lease, interpreting personal finances, and building software such as a game or business app.¹ The company says the models are designed for expert domains including law, finance and health, as well as coding tasks that require sustained iteration across requirements, tests and revisions.¹²

The model card lists input support for text, image, video and audio, with text output only. It also lists a one million-token inference context length, which matters for document-heavy and tool-heavy agent workflows that need to ingest large files, logs, codebases or conversation histories.³

Intended uses include general assistance, reasoning, summarization, document and data analysis, coding, software engineering, multi-step workflows, domain-specific knowledge work, multimodal understanding, and research and development.³

Meta says its evaluations cover personal and professional agents, memory, coding, domain expertise and general intelligence.² The model card compares Muse Wonder and Muse Wonder Full with OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Claude Opus 5.5, generally at maximum reasoning effort, while noting that external leaderboard results and third-party harnesses can make comparisons imperfect.³

SiliconANGLE highlighted those caveats, noting that some coding results were not controlled apples-to-apples comparisons because different agent products or reasoning settings were used.⁶

TechCrunch framed the release as Meta’s first attempt at a frontier model. It reported that Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API, while Muse Wonder Full is not served through the API and is limited to Muse Code subscribers.⁵

Active Memory is the product bet

The most strategically important feature is Active Memory. Meta defines it as the model’s ability to carry important context about a user across conversations and retrieve relevant details when a new task requires them.²

Axios described it as a way for an AI system to retain important user information across days, weeks or months so users do not have to repeatedly explain their preferences, goals and constraints.⁴

For platform teams, that moves memory from a chat feature into a core layer of agent infrastructure. Persistent memory can make agents more useful in workflows such as recurring planning, enterprise support, personal finance, software maintenance or health-related document review.

It can also create new product obligations: memory permissions, deletion controls, provenance, audit logs, cross-application data boundaries and policies for when a model should not use remembered information.

Meta has said Muse Wonder was trained to avoid sharing private information without permission, and its research post says personal agents may handle messages, calendars and private documents.¹² But the launch materials do not yet provide a detailed privacy architecture for Active Memory, such as how memories are stored, scoped, retained, edited, deleted or separated across consumer, developer and enterprise contexts.

That gap is likely to matter for teams evaluating whether to embed the models in regulated or workplace products.

Active Focus targets long-running agents

Meta’s second major feature, Active Focus, is designed to keep an agent working toward a goal when tasks become difficult. The company says the model can try alternative approaches, work through blockers, adapt to new information, ask for clarification and check in before taking important actions.²

Meta says it evaluates focus through FocusBench and sustained execution through SWE-Marathon, a benchmark for long-running software projects.²

That design choice reflects a broader platform problem: agents often fail not because they cannot answer a single prompt, but because they drift, loop, lose state, misuse tools or compound small errors over long tasks. WIRED noted that Active Memory and Active Focus are meant to keep agents locked onto a goal and reduce confusion during complex work such as research or large software projects.⁷

For product teams, the relevant question is how much of Active Focus is model behavior versus system scaffolding. A long-running agent usually needs external planning, state management, tool permissioning, rollback, observability and human approvals.

Meta’s model card recommends deploying Muse Wonder as part of a broader AI system with additional guardrails, including human-in-the-loop confirmation for irreversible actions.³

Safety disclosures are partial

Meta says it trained the Muse Wonder family to refuse harmful requests, protect private information and support defensive cybersecurity work without enabling misuse.¹ The company says a full report with additional safety evaluations will be available soon.¹² Until then, the model card is the main public technical safety document.

That card says Muse Wonder Full, the strongest variant, was assessed under Meta’s Superintelligence Scaling Framework across cybersecurity, chemical and biological risks, and loss of control. It says Meta does not rule out the unmitigated model reaching a high-risk threshold in the chemical and biological domain, so the company applied mitigations including safety post-training and system-level classifiers.³

With mitigations and system safeguards in place, Meta says residual risk is “moderate or lower” across the evaluated domains.³

The model card also reports strong refusal performance on high-risk biological and chemical prompts and says Muse Wonder Full had a 0% harmful-action rate on an agentic misalignment evaluation.³ SiliconANGLE reported that the cyber results include an ExploitGym caveat: although the model captured 92 flags, only 21 used the intended vulnerability path, which is the figure comparable with the published ranking.⁶

WIRED raised a separate issue: Meta has not disclosed whether Muse Wonder-powered agents misbehaved during testing or what mitigations it used for agent-specific failure modes beyond the general safety framing.⁷

The model card also states that testing cannot cover all scenarios, that the model may still produce inaccurate, biased or objectionable responses, and that it may make errors in multi-step reasoning, especially in novel settings.³

The unresolved platform questions

Muse Wonder gives Meta a stronger answer to a question that has followed its AI strategy: whether its advantage is distribution, research or infrastructure. The launch suggests Meta wants all three, but the new emphasis is infrastructure for persistent, capable agents.

For AI product and platform teams, three issues should shape evaluation.

First, benchmark claims need independent validation, especially for agent workflows where tool harnesses, prompts, latency and recovery loops affect outcomes. Second, Active Memory requires clear product controls before use in sensitive consumer or enterprise settings. Third, Active Focus may reduce some long-horizon failures, but it does not replace agent governance, observability and approval systems.

Meta’s own materials describe Muse Wonder as a step toward personal superintelligence.¹² The more immediate implication is narrower and more practical: Meta is building models for agents that persist, remember and act.

The company has now made that strategy public. The documentation needed to judge its privacy and safety posture is still catching up.