On Monday, October 12, Meta unveiled two new artificial intelligence models, Muse Wonder and Muse Wonder Full, which it describes as its most powerful models. According to the report published by Journal du Coin, Mark Zuckerberg’s group thus aims to compete more directly with OpenAI and Anthropic. The launch comes fourteen months after the start of an overhaul of its AI activities.
As is often the case with this type of announcement, three levels need to be distinguished: what the company claims, what has actually been shown, and what remains to be established through independent evaluations or through use.
What is announced: models for long tasks
The stated ambition goes beyond a one-off answer to a question. Meta cites concrete examples: understanding a lease, organising a claim with an insurer or building an application. To achieve this, the models must be able to work with several documents, follow instructions and chain together the steps needed to complete a task.
Two features underpin this promise:
- Active Memory is meant to retain important information from one conversation to the next and retrieve it when it becomes useful. The aim is to spare users from having to systematically restate their preferences, constraints or the details of a project.
- Active Focus refers to the model’s ability to pursue a task despite difficulties: trying other approaches, correcting its mistakes and asking for clarification when necessary.
Meta also says it has trained its models to consult the user before important actions. Software development plays a central role: Muse Wonder and Wonder Full are meant to follow a project’s requirements, modify code and repeat testing phases.
What is demonstrated: in-house evaluations
On the question of performance, caution is in order. Meta claims top-tier results, but, as Journal du Coin points out, the results published at the launch remain evaluations presented by the company itself. They are not enough to establish general superiority over its competitors.
The group also released a promotional and explanatory video on its X account. A demonstration of this kind illustrates the product’s intent; it does not constitute a reproducible measure of its capabilities.
In other words, what is established at this stage is the launch itself and the availability of Muse Wonder through certain channels. Superiority over OpenAI and Anthropic, on the other hand, is a company claim that remains to be confirmed. As for the ability to handle an insurance case or a software project from start to finish without error, this is a prospect, not a fact observed in the available sources.
What remains to be verified: security and privacy
An assistant that works from messages, calendars and personal documents necessarily has access to sensitive information. Persistent memory, presented as an asset, heightens this concern: the more the model retains, the more central the question of what it does with that data becomes.
Meta says it has trained these models not to share private data without authorisation and to resist malicious instructions. These, again, are commitments made by the company. A report including additional security evaluations has yet to be published. Until it is available, it is not possible to judge to what extent these safeguards hold up in real-world use.
Actual availability: paid and gradual
On access, the information is more precise:
- Muse Wonder is available via the Meta Model API and via Muse Code, the group’s programming tool. The announced price is 4 dollars per million input tokens and 20 dollars per million output tokens, tokens being the units of text processed or generated by the model.
- Muse Wonder Full is reserved for Muse Code subscribers.
- Muse Wonder is also due to join the Muse personal assistant soon, with an initial rollout to subscribers. Meta has not announced a specific date for free access.
For users in Switzerland and Europe, the source consulted provides no information: it does not specify whether the API and Muse Code are accessible from these countries, nor whether the integration into the Muse assistant will be offered there, nor on what timetable. In the absence of data, it is not possible to state that these models are, or will be, available to the Swiss or European public. Those interested will need to refer to the access conditions communicated by Meta.
What this changes in practice
For now, the first to be affected are developers, who can test Muse Wonder via the API or Muse Code and judge its programming capabilities first-hand. For the general public, an assistant capable of handling a lease or a claim remains a promise whose rollout has been announced, but with no date for free access.
Strategically, this direction opens up an additional monetisation route for Meta’s AI investments. It also shifts the competitive landscape: beyond performance scores, the value of an autonomous assistant will be measured by its reliability on long tasks and by the way it handles the data entrusted to it. Two points that only independent evaluations and the announced security report will make it possible to assess.
