Meta believes Muse Wonder and Muse Wonder Full models have touched the frontier

The key to these advances is Active Memory and Active Focus, and the company believes it leads the Personal Superintelligence Index evaluations.

This may mark a milestone for Meta, after spending tens of billions of dollars on its artificial intelligence (AI) efforts. The company unveiled its new Muse models, Muse Wonder and Muse Wonder Full, on Tuesday, saying they mark “a major leap toward personal superintelligence”. Meta is benchmarking the new models for a comparative illustration of how they stack up against Anthropic’s Claude Opus 5.5, OpenAI’s GPT-6 Astra, Google Gemini 4 Argon and their own Muse Spark models.

Key to advancements are Active Memory and Active Focus capabilities, and Meta believes it leads Personal Superintelligence Index evaluations (Official image)

The comparison with Muse Spark models is important because, when Meta introduced them earlier, it also outlined a rigorous scaling approach. “Today’s release builds on that work across pretraining and reinforcement learning, delivering our most capable models yet,” Meta said in a statement.

The key to these advances is Active Memory and Active Focus, and the company believes it leads the Personal Superintelligence Index evaluations. With all models set to max reasoning effort, the Muse Wonder Full scored 59.1 in the Personal Superintelligence Index, while GPT-6 Astra scored 54.5 and Claude Opus 5.5 scored 55.2, placing ahead of Muse Wonder (54.3).

There is also a pitch for being adept at handling tasks in law, finance, and health domains. On the Harvey LAB-AA benchmark, which tests a model’s capabilities as a legal services agent, Muse Wonder Full scored 13.1, while Muse Wonder scored 10. The closest competitor is GPT-6 Astra, with an 8.6 score. The Vals Finance Age benchmark for financial analyst research sees Muse Wonder Full score 59.1, followed by Claude Fable 5.1 (58.9), Claude Opus 5.5 (58.6) and Muse Wonder (56.6). In Healthbench Pro, Claude Opus 5.5 leads with a 65.6 score, followed by GPT-6 Astra (65.3) and Muse Wonder Full (62.3).

“Trained at our gigawatt-scale Prometheus data center, Muse Wonder and Muse Wonder Full bring together the work of teams across Meta to deliver frontier capabilities for complex tasks at work and in everyday life — another step toward bringing personal superintelligence to the billions of people who use our products,” they say. This ties in with plans for upcoming personal agent hardware.

There have been calls for a dedicated AI device, but no one knows what form or use will catch on with users. Meta showed its own, the Muse Charm, at its Connect keynote last month. It is envisioned as a wearable with a lanyard that lets users talk to the Muse-powered agent without needing a phone. Mark Zuckerberg, Meta’s chief executive, had said only a few had been built so far. Meta plans to ship the Charm “in time for the holidays in December”.

Muse personal agents are already available on Mac computing devices, smartphones, the web, and inside WhatsApp. Users can connect third-party apps, including calendars, email clients, and work apps, to perform agentic tasks such as filling web forms, checking for pricing changes on e-commerce platforms for an item you wish to buy, reserving tickets, managing calendars, and curating a fitness plan. Meta insists the agent cannot read app logins, for instance, underscoring its privacy pitch.