An AI agent that reads a lease, chases an insurance claim and untangles someone's finances is the pitch behind Meta's new Muse Wonder models. Meta claims its flagship now beats rival frontier models at legal document work.
Muse Wonder is available to developers through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Muse Wonder will soon power Meta's Muse personal agent, starting with subscribers.
The standout result is legal. Harvey's Legal Agent Benchmark tests whether a model can work from source documents and follow detailed instructions to produce legal work products.
In Meta's testing, Muse Wonder Full scored 13.1 on the benchmark. GPT-6 Astra scored 8.6, Claude Fable 5.1 scored 6.4 and Claude Opus 5.5 scored 4.2, according to Meta's research blog.
Finance was a closer contest. On Vals Finance Agent V2, which measures financial analyst research, Muse Wonder Full edged Claude Fable 5.1 by 59.1 to 58.9.
The model also led on AutomationBench-AA, which tests end-to-end business workflows, scoring 45.4 against 41.6 for GPT-6 Astra. Health was a different story. On HealthBench Pro, Muse Wonder Full's 62.3 trailed Claude Opus 5.5 at 65.6.
The agent is built to remember. Meta's Active Memory feature carries context about a user across conversations and brings back details when they become relevant to a task.
On Meta's MemoryBench, Muse Wonder scored 85.6, ahead of 81.5 for Claude Opus 5.5. A second feature, Active Focus, keeps the model working on a goal, trying other approaches when stuck and checking in before important actions.
Meta paired the memory pitch with privacy claims. It said a personal agent may handle messages, calendars and private documents, and that it trained the models to avoid sharing private information without permission.
In Meta's tests of mishandling personal information, Muse Wonder Full had a violation rate of 2.2%. That compares with 4.6% for GPT-6 Astra and 6.8% for Claude Fable 5.1.
Meta also tested resistance to manipulation. Attacks designed to make the models execute unauthorised code succeeded against Muse Wonder Full 10.7% of the time, compared with 19.7% for GPT-6 Astra.
Muse Wonder did not top every chart. Claude Fable 5.1 refused cyber misuse requests slightly more often, at 97.5% against 96.9%. Claude Opus 5.5 led on several coding tests and on the Artificial Analysis Index.
"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," Meta said.
Meta said a full report with additional safety evaluations will be available soon. The models were trained at its Prometheus data centre and build on its earlier Muse Spark model.