Meta is now comparing its artificial intelligence models to OpenAI’s GPT-6 Astra, as well as Anthropic’s Claude Fable 5.1 and Claude Opus 5.5. On this occasion, the group is introducing Muse Wonder and Muse Wonder Full, a new generation it describes as ‘a major leap towards personal superintelligence’. Until now, Meta relied on Muse Spark, a lighter model designed for fast responses, which notably powers its Muse agent.
Models Designed for Long and Specialized Tasks
Muse Wonder and Muse Wonder Full target agentic tasks, meaning those where the model must chain several actions to achieve a goal, whether personal or professional. Meta highlights expert domains such as law, finance, or health, from reading a lease to dealing with an insurer. The group also claims a state-of-the-art level in programming. Both models accept text, images, video, and audio as input, with a context window of one million tokens. A token is a unit of text that the model processes, and the context window refers to the amount of information it can consider at once.
According to Meta’s research blog, the base model, larger and more robust, makes the range more comfortable with tasks that were not specifically trained for, an asset for a personal agent facing unexpected requests. Meta also unveils two new capabilities. Active Memory involves retaining important information about the user from one conversation to the next and recalling it at the right moment. Active Focus refers to the model’s ability to pursue a goal despite obstacles: it changes approach when a path fails, recovers from its errors, and requests validation before any significant action.
In its demonstrations, Meta shows Muse Wonder building a 3D cooking game from images, or organizing a wedding seating plan according to a series of constraints.
Safeguards for Access to Private Data
A personal agent can handle sensitive data, such as messages, calendars, or private documents. Meta states it has trained its models not to share personal information without authorization and to resist malicious instructions. The issue is all the more sensitive as the Muse agent is among the AI applications most hungry for personal data, according to the latest ranking by Surfshark. Meta also announces its strongest safeguards to date against misuse related to chemical, biological, radiological, or nuclear weapons, as well as cyberattacks.
Access Reserved for Developers, Awaiting the Muse Agent
Muse Wonder is available via the Meta Model API and in Muse Code, Meta’s programming agent. Its price, mentioned by TechCrunch, is $4 per million input tokens and $20 per million output tokens, compared to $1.25 and $4.25 for Muse Spark 1.2 at the launch of Muse Code. Muse Wonder Full, the most powerful version, is not offered via the API and remains reserved for Muse Code subscribers.
Muse Wonder is set to power the Muse agent ‘soon’, starting with its subscribers, with no specific date. Meta also plans to use it for artifact generation and to develop specialized models for Muse users. The Muse agent, available in the United States since September, is still not offered in France or the rest of the European Union, and Meta has not communicated any deployment date. For developers, however, the announcement mentions no geographical restrictions regarding the Meta Model API and Muse Code.
