Meta has released the most capable AI models in its new Muse family, and the company says they stand alongside rivals’ best. Muse Wonder and Muse Wonder Full, announced by Meta Superintelligence Labs, are the company’s first attempt at what the industry calls a frontier model, larger and more capable than the lightweight Muse Spark that preceded them, according to TechCrunch.

Meta’s blog post calls the models “a major leap toward personal superintelligence.” The claims behind that, though, come from Meta’s own testing, and the sources reviewed here found no independent verification of the full benchmark picture.

What is new

Meta says the models deliver “state-of-the-art performance on personal and professional agentic tasks,” aimed at expert fields such as law, finance and health, as well as coding for building games and apps. The pitch rests on two features. Active Memory carries relevant context across conversations, which TechCrunch describes as retaining and synthesising what users share over time for long-running personal assistant tasks. Active Focus is meant to keep the model working toward a goal when a task gets hard, including trying other approaches, recovering from errors and asking for confirmation before important actions, according to Meta’s announcement.

“Remember what’s important to you over days, weeks, months and beyond,” said Kavya Kopparapu, a research scientist at Meta, in the launch video, describing the memory work, as TechCrunch reports.

Who can use them, and at what price

Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, which TechCrunch says is roughly on par with OpenAI’s GPT-6 Sol and Anthropic’s Claude Opus 5.5. Muse Wonder Full is limited to subscribers of Muse Code, Meta’s coding product, and is not served through the API. Both will power the Muse agent, and Meta says Muse Wonder will soon power experiences in the Muse app, starting with subscribers.

Kingy.ai, an analysis site that cites Meta’s pricing documents, lists Muse Code plans from $5 to $200 a month, with Wonder Full included only in the $50 and $200 tiers. It also notes that subscriptions default to a mode that lets prompts and code be used to train future Meta models, while direct API use does not. These details were not in the Meta announcement page reviewed here and should be checked against Meta’s documentation.

How it measures up, according to Meta

Meta says both models score comparably to Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra on its benchmarks, and higher on MemoryBench, a test run by Meta, than Fable, Opus or Astra. The gap on other tests is mixed. Kingy.ai, citing Meta’s scorecard, reports that Wonder Full leads some coding and professional-task tests while Opus 5.5 leads others, such as Terminal-Bench and ProgramBench.

Independent data points to a narrower picture. Kingy.ai cites Artificial Analysis, a third-party tracker, as scoring Opus 5.5 at 57.6, Wonder Full at 55.8, Fable 5.1 at 53.4, GPT-6 Astra at 52.7 and Muse Wonder at 52.4 on its index. By that measure, Wonder Full sits second behind Opus, and the standard Wonder model is behind both Fable and Astra. Gagadget, which covered the release, noted that no one outside Meta had checked its performance claims, so the “frontier” label remains unproven.

The strategic context

The launch is the first frontier-level release since Alexandr Wang restructured Meta’s AI efforts in August 2025, TechCrunch notes. Since then Meta has shipped Muse Spark and the Muse agent but had avoided the kind of state-of-the-art releases associated with OpenAI and Anthropic. TechCrunch says the models were trained at Meta’s Prometheus data centre in central Ohio, which came online earlier this year.

Whether the weights are public is the open question. TechCrunch’s report does not say, and Gagadget states that neither model has public weights, unlike Meta’s earlier Llama models. Meta did release a model called Muse Glimmer as a freely downloadable model in August, according to The Nauti. Gagadget also notes that EU rules for general-purpose AI became enforceable in August 2026, and that Meta has not said how Muse Wonder complies, advice it frames for European developers.

What to watch

Meta says a full report with additional safety evaluations is “available soon.” Outside testing of Wonder and Wonder Full, any public move on weights, and developer uptake at API prices level with Opus 5.5 will show whether the benchmark story holds up in practice.