Meta launches Muse Wonder and Muse Wonder Full with new safeguards and pricing

On October 12, 2026 Meta announced the release of two new large‑language models, Muse Wonder and Muse Wonder Full, describing them as the company’s most performant models to date (Source 1).

Muse Wonder is immediately available to developers through Meta’s Model API and the Muse Code platform, while Muse Wonder Full is restricted to Muse Code subscribers only (Source 1).

Core capabilities for personal and professional agents

Both models target agentic tasks across a range of industries. They are marketed for insurance claims handling, lease interpretation, financial analysis and health‑care workflows (Source 1).

Technically, the models incorporate two novel mechanisms: “Active Memory” retains relevant conversational context, and “Active Focus” enables the system to persist through difficulties, try alternative approaches and request clarification when needed (Source 1).

Safety and misuse prevention

Meta claims Muse Wonder includes its strongest safeguards ever, refusing to share private information without explicit permission and declining to assist with dangerous actions (Source 2).

A full security assessment using Meta’s “Superintelligence Scaling Framework” is slated for publication in the coming weeks, according to the company’s own statements (Source 1).

Pricing structure for developers and enterprises

Access to Muse Wonder via the Muse app requires a monthly subscription of $20 for basic usage or $100 for higher‑volume tiers (Source 2).

Muse Wonder Full, the more powerful variant, will be offered under Muse Code with subscription options of $50 per month for standard access or $200 for premium usage (Source 2).

• $20/month basic Muse Wonder
• $100/month high‑volume Muse Wonder
• $50/month standard Muse Wonder Full
• $200/month premium Muse Wonder Full

The underlying model, codenamed “Watermelon,” was trained specifically to handle unexpected obstacles and task blockages, experimenting with multiple strategies before looping back to the user for guidance (Source 2).

Following the launch of the Muse app a month earlier, it quickly rose to the top of Apple’s App Store rankings, despite the base application remaining free; the revenue driver is the tiered subscription model (Source 2).

The introduction of Muse Wonder and Muse Wonder Full brings a clear framework for evaluating how agentic AI can be integrated into existing workflows. By offering a tiered subscription model, developers can benchmark performance against cost, allowing a systematic comparison with prior generations of models. This structure also creates a measurable baseline for assessing whether the added capabilities—such as the ability to retain context over longer interactions—translate into tangible productivity gains. In practice, organisations can set up controlled pilots that compare task completion times and error rates before and after deployment, thereby quantifying the operational impact of the new mechanisms without relying on anecdotal evidence.

From a safety perspective, the claimed “strongest safeguards ever” introduce a set of verification checkpoints that must be rigorously tested in real‑world settings. The refusal to share private information without explicit permission implies the presence of an intent‑recognition layer that must be validated against a wide spectrum of data‑handling scenarios. Moreover, the model’s refusal to assist with dangerous actions requires a robust classification system for identifying risky requests. Systematic auditing—through red‑team exercises and continuous monitoring—will be essential to confirm that these safeguards operate consistently and do not produce false positives that hinder legitimate use cases.

The novel mechanisms dubbed “Active Memory” and “Active Focus” raise important questions about the limits of autonomous problem solving. While the ability to retain relevant conversational context can reduce the need for repeated user input, it also introduces the risk of propagating outdated or incorrect assumptions if the model fails to recognize a shift in the task’s parameters. Similarly, the persistence feature that tries alternative approaches before seeking clarification may lead to longer processing cycles, potentially impacting latency‑sensitive applications. Evaluating these trade‑offs requires detailed logging of interaction paths, enabling developers to measure how often the model cycles through strategies versus how frequently it escalates to human intervention.

Verification of the “Superintelligence Scaling Framework” assessment will be a critical step before organisations can place full confidence in the models’ reliability. Even though the framework is slated for publication, its practical utility hinges on transparent methodology, reproducibility of results, and independent peer review. Until such validation is publicly available, stakeholders should treat the security claims as provisional, incorporating additional layers of internal testing—such as fuzzing inputs and stress‑testing under peak loads—to uncover any latent vulnerabilities that might not be captured by the initial assessment.

The pricing tiers, while straightforward, also impose practical constraints on adoption patterns. Lower‑cost tiers may limit the volume of API calls, which could restrict the model’s usefulness for large‑scale batch processing or high‑frequency user interactions. Conversely, premium tiers provide the computational headroom needed for intensive tasks but increase financial exposure, especially for smaller enterprises. Decision‑makers must therefore conduct cost‑benefit analyses that factor in not only subscription fees but also ancillary expenses such as monitoring infrastructure, compliance audits, and potential downtime arising from safety overrides.

Finally, the broader consequences of deploying these agentic models revolve around regulatory compliance and risk management. The built‑in privacy safeguards can simplify adherence to data‑protection statutes, yet organisations must still document how consent is obtained and verified within their specific workflows. The models’ ability to request clarification introduces a new interaction pattern that may affect user experience design, requiring UI adjustments to accommodate back‑and‑forth dialogue. In the long term, the presence of robust safety controls and transparent pricing may encourage wider industry adoption, but only if the promised performance and security benefits are demonstrably upheld through continuous, independent verification.

For English‑speaking organisations, the launch means immediate access to a high‑performance agentic AI that can be embedded in internal tools via API, with clear pricing tiers and robust safety controls that reduce legal exposure and simplify compliance with data‑privacy regulations.