Meta releases Muse Wonder, which it calls its most capable model, for personal and professional agents

Meta released Muse Wonder and Muse Wonder Full on Monday 12 October 2026, describing them in a blog post as "our most capable models" and saying they "deliver state-of-the-art performance on personal and professional agentic tasks". TechCrunch described the release as "the company's first attempt at a frontier model"; Meta's own post does not use the word "first" about the models.

Key takeaways

• Meta says Muse Wonder is built around what it calls "Active Memory" and "Active Focus", for personal and professional agents.
• Muse Wonder is on the Meta Model API at $4.00 per million input tokens and $20.00 per million output tokens, according to Meta's pricing page. Muse Wonder Full is only in Muse Code, on two subscription plans.
• Meta's model card sets out its risk assessment, but the blog says "a full report with additional safety evaluations will be available soon".

What Meta says it is for

Meta calls the pair "a major leap toward personal superintelligence", a phrase Meta uses throughout the post. The post rests on two ideas:

• Active Memory: "carrying important context about you across conversations and actively retrieving the details relevant to the task at hand".
• Active Focus: "trying alternative approaches, working through blockers, and adapting to new information". Meta says the model "asks for clarification when needed and checks in before taking important actions".

Meta also points to work in law, finance and health. The model card lists text, image, video and audio input, text output, and a context length of "1M tokens at inference". Meta's developer documentation says Muse Wonder "is in preview".

Where it runs and what it costs

• Training: Meta says both models were trained at its "gigawatt-scale Prometheus data center". TechCrunch says that site is in central Ohio.
• Availability: "Muse Wonder is available today through the Meta Model API and Muse Code, and Muse Wonder Full is available today for Muse Code subscribers", Meta's post says. Its documentation adds that Wonder Full is "available only in Muse Code with a subscription", on the Power Usage and Maximum Usage plans, and is not offered on the pay-as-you-go API.
• Price: Meta's pricing page lists $4.00 per million input tokens, $20.00 per million output tokens and $0.20 per million cached input tokens. TechCrunch reported the same input and output prices.
• Training on your data: in Muse Code, Meta says subscriptions default to a "Contributor mode" that gives more usage "in exchange for letting Meta use your content to train future Meta models", and that users can turn it off. In Standard mode, Meta says content "is not used to train Meta models".

UK and EU: Meta's Geographic Use Policy for its API, updated 12 October, lists no country restriction specific to Muse Wonder; no UK or EU country is on its restricted list. It also says the services "are not available in all jurisdictions". None of the Meta pages we read states whether Muse Wonder or Muse Code can be used from the UK or EU, and we have not tested access.

Benchmarks and the "benchmaxxing" question

Meta publishes comparisons with rival models in its post and model card. In Meta's main comparison table, a Muse Wonder model has the top score on 7 of 14 rows and a rival model on the other 7; in several rows the margin is a fraction of a point. The model card says competitor runs it generated itself "remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models".

The more unusual section of the post addresses the complaint known as "benchmaxxing". "Two models often have similar benchmark scores, yet one seems to handle a much wider range of unfamiliar tasks", Meta writes. It compared Muse Spark 1.3 and Muse Wonder Full on tasks it targeted in post-training and tasks it did not, at several reasoning-effort settings. Meta says Wonder Full "performs even better on unfamiliar tasks than the Muse Spark 1.3 trend would predict", and attributes that to "larger, stronger base models". The post does not name the tasks in either set.

Safety

Meta says Muse Wonder "offers our strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse, while preserving access to frontier cyber capabilities for defensive use".

The model card gives Meta's risk assessment now. It says a model "reaches the high risk threshold when its deployment could substantially contribute to a threat scenario we have identified as severe", and: "Without mitigations, we do not rule out the model reaching this threshold in the Chemical & Biological domain", so Meta applied safety post-training and system-level classifiers. With those in place, Meta says it reached "moderate or lower" across all domains, "which supports our decision to deploy". The card also says the model "is not intended to be used by individuals under the age of 18", and recommends "human-in-the-loop confirmation for irreversible actions" in agentic use.

Earlier coverage

We have covered Meta's Muse before, including WIRED's report that instructions extracted from the Muse agent appear to describe "a page for every person in the user's life", context that Meta said Muse gathers from public information and what users have chosen to share.

Why it matters

On our reading, the release puts Meta's personal-agent pitch onto a paid, higher-priced developer model: on Meta's own price list, Muse Wonder's input tokens cost about three times Muse Spark's Standard-mode price ($4.00 against $1.25) and its output tokens nearly five times ($20.00 against $4.25). The pitch rests on remembering the user and persisting with long tasks, which on our reading are also the features that matter most for privacy in a personal agent. The performance claims are Meta's own, the model is in preview, and the fuller safety report has not yet been published.