Meta unveils models it says are frontier-class
Muse Wonder and Muse Wonder Full are aimed at agents that can remember context and complete extended tasks.
Meta has launched two new artificial-intelligence models as it tries to close the gap with OpenAI and Anthropic. The company describes Muse Wonder and Muse Wonder Full as its most capable models yet, but independent testing of the performance claims is still limited.
Meta announced Muse Wonder and Muse Wonder Full on Monday, presenting them as a major step in its effort to compete at the leading edge of artificial intelligence. The models are designed for tasks that require more than a single answer, including coding, handling documents and working through professional problems over time.
Muse Wonder is available to developers through Meta’s Model API and to users of Muse Code, while Muse Wonder Full is initially available to Muse Code subscribers. Meta said the Wonder family would also support its personal AI agent, Muse, while wider access was expected to begin with subscribers.
The company is emphasising two features called Active Memory and Active Focus. Active Memory is intended to retain useful information across conversations, while Active Focus is designed to let the system continue working, recover from errors and ask for clarification before taking an important action.
Meta says the models deliver state-of-the-art performance on personal and professional agentic tasks, including work involving law, finance and health. These are company claims, and the full safety evaluation promised by Meta had not yet been published when the models were launched.
Independent coverage of the release reported that Meta’s own benchmark results placed the models alongside or ahead of named rival systems on selected tests. Such comparisons can indicate where a model performs well, but they do not establish that it is broadly better: results depend on the tasks, prompts, evaluation methods and whether outside researchers can reproduce them.
The launch also raises practical questions about privacy and control. An agent that can remember personal information and act across messages, calendars and documents may be more useful than a chatbot, but it also has more opportunities to make an unwanted disclosure or an incorrect decision. Meta says its systems are trained to seek permission for sensitive actions and to reduce harmful behaviour, while promising more detailed safety information later.
For Britain, the significance is less about the label ‘frontier’ than about how quickly such agents move into ordinary services and workplaces. The immediate test will be whether Meta’s claimed capabilities survive independent evaluation and whether users can understand, limit and audit what the systems remember and do on their behalf.
One story, several perspectives
What is established
- Meta launched two models and says they are intended for extended personal and professional tasks.
- The models can be accessed through Meta products and services, with some access initially limited to developers or subscribers.
- Meta’s performance claims are based partly on its own evaluations, while broader independent testing remains limited.
Left
Arguments The rapid deployment of powerful agents should be slowed until privacy, labour-market effects, energy use and accountability are better understood. Concentrating advanced systems inside a small number of technology companies risks giving private firms too much influence over everyday information and work.
Values Collective oversight, worker protection, privacy, equality and democratic control over infrastructure.
Consequences This approach would favour stronger regulation, independent audits and limits on high-impact uses, even if product releases and investment moved more slowly.
Centre
Arguments Innovation should continue, but deployment should be tied to transparent evaluations, meaningful user consent, clear audit trails and enforceable remedies when an agent causes harm. The question is not whether AI can be stopped, but whether its risks can be governed as its capabilities grow.
Values Evidence, proportionality, consumer protection and accountable innovation.
Consequences It would support controlled rollout with independent testing and sector-specific rules rather than either a blanket pause or unrestricted release.
Right
Arguments Competition between private companies is the main engine of progress, and Meta’s challenge to OpenAI and Anthropic could reduce dependence on a small number of incumbents. Users and developers should be able to choose between systems, while regulation should target clear harms rather than speculative risks.
Values Innovation, enterprise, consumer choice and limited state intervention.
Consequences This view expects faster adoption and lower costs, but accepts that some risks will have to be managed after products reach the market.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Priya Shah — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The article separates Meta’s confirmed product launch and stated capabilities from claims that depend on company benchmarks. No core factual error was found, and the limits of independent verification are stated explicitly.
- confirmed Meta launched Muse Wonder and Muse Wonder Full on 12 October 2026. — Confirmed in Meta’s launch announcement and same-day independent coverage. source
- confirmed Muse Wonder is available through Meta’s Model API and Muse Code, while Muse Wonder Full is initially available to Muse Code subscribers. — Stated in Meta’s launch announcement. source
- confirmed Meta says the models are designed for extended agentic tasks and include Active Memory and Active Focus. — These are product descriptions in Meta’s announcement, presented as company claims. source
- confirmed Independent reporting described Meta’s benchmark comparisons with rival systems. — TechCrunch reported the comparison while noting that the results came from Meta’s testing. source
- uncertain Independent testing has not yet established that the models are broadly better than rivals. — The launch coverage identified the need for independent evaluation; the statement is an editorial assessment of the available evidence rather than a measured product result. source
Editor's note
The launch, availability and product claims are confirmed by Meta. Benchmark comparisons and the frontier-level description remain company claims pending broader independent testing; Meta also said a fuller safety report would follow.Sources
More on this in other media
- iNews — „meta artificial intelligence”
- BBC News — „meta artificial intelligence”
- The Guardian — „meta artificial intelligence”