On Tuesday, Meta introduced two additions to its Muse model lineup — Muse Wonder and Muse Wonder Full — with the company characterising them as representing a substantial advance toward what it calls personal superintelligence. Performance comparisons against Anthropic's Claude Opus 5.5, OpenAI's GPT-6 Astra, Google Gemini 4 Argon, and Meta's own Muse Spark models are currently being run to illustrate where the new releases stand relative to those systems.
That comparison to the Muse Spark models holds some analytical grounding, given how Meta's rollout paired the product announcement with a published breakdown of its scaling approach — a framework the company's own statement attributed to advances in pretraining and reinforcement learning, positioning the resulting lineup as its most capable to date.
Two proprietary technologies — Active Memory and Active Focus — are credited as the driving force behind those developments, with the firm asserting that its standing atop the Personal Superintelligence Index evaluations places it ahead of rivals.
Running on maximum reasoning settings, Muse Wonder Full posted a Personal Superintelligence Index reading of 59.1, with Claude Opus 5.5 returning 55.2 and GPT-6 Astra 54.5, both finishing above the 54.3 recorded by Muse Wonder. Capabilities spanning legal, financial, and healthcare domains are also part of the offering put forward.
Testing Muse Wonder Full against the Harvey LAB-AA benchmark — a framework designed to evaluate how well a model performs as a legal services agent — placed it at 13.1, three points above the 10.0 recorded by Muse Wonder on the same assessment. Running GPT-6 Astra at 8.6 placed it nearest on the leaderboard, trailing only the top-ranked model.
On the Vals Finance Age benchmark measuring financial analyst research, Muse Wonder Full posted a leading score of 59.1, with Claude Fable 5.1 recording 58.9, Claude Opus 5.5 at 58.6, and Muse Wonder rounding out the group at 56.6.
On the Healthbench Pro assessment, a three-way contest at the top placed Claude Opus 5.5 narrowly ahead with 65.6 points, while GPT-6 Astra recorded 65.3 and Muse Wonder Full came in at 62.3. On the Artificial Analysis Intelligence Index benchmark, a score of 58 placed Claude Opus 5.5 at the top of the table, with Muse Wonder Full posting 56 to land one position behind.
Outperforming Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon — each of which posted a score of 53 — as well as Muse Wonder at 52, the new models drew on work spanning multiple internal teams. The company noted in a statement that both Muse Wonder and Muse Wonder Full were developed at its gigawatt-scale Prometheus data centre and are intended to handle demanding tasks in professional and daily-life contexts as part of a broader effort to extend what the company described as personal superintelligence to the large user base across its platforms.
Those development directions feed into a broader roadmap centred on dedicated hardware for personal AI agents. During its Connect keynote held roughly a month prior, Meta introduced a wearable device under the name Muse Charm. Designed as a lanyard-worn device, it gives wearers direct access to the Muse-powered agent through voice alone, with no phone required.
Zuckerberg noted that construction of the units remained at an early stage, with only a handful completed. According to Meta's stated timeline, the Charm device is set to reach consumers before the end of December, while Muse personal agents have already been rolled out across Mac computers, phones, web browsers, and WhatsApp.
Third-party apps — calendars, email clients, and workplace tools among them — can be linked to the platform so it handles tasks on the user's behalf, from submitting web forms and tracking price shifts on shopping sites to booking tickets, organising schedules, and putting together a personalised fitness routine. According to the company, the agent lacks the technical ability to access application login credentials — a position Meta has put forward as part of its broader data-handling stance.
