Meta’s Muse Wonder: Its First Frontier Model, Benchmarked

Meta Superintelligence Labs released two models on 12 October, Muse Wonder and Muse Wonder Full, which Meta describes as its most capable models for personal and professional agentic tasks and “the next step on our scaling ladder”. They are the lab’s first models aimed squarely at the frontier rather than at speed: Muse Spark, which opened the series in April, was built to be light and fast. Both were trained at Prometheus, Meta’s gigawatt-scale data center in Ohio.

The two are not sold the same way. Muse Wonder is available now through the Meta Model API and Muse Code, at $4 per million input tokens and $20 per million output tokens, TechCrunch reported. Muse Wonder Full, the larger and stronger of the pair, is only available to Muse Code subscribers and is not served through the API. Meta says Wonder will “soon” power its Muse personal agent app, starting with paying subscribers. According to the model card, both take text, images, video and audio as input, produce text, and have a 1-million-token context window.

Memory as the selling point

Meta is leading with two behaviours rather than a single score. “Active Memory” is meant to keep the context that matters about a user across conversations and pull the relevant parts back into the task at hand. “Active Focus” describes an agent that keeps working at a hard goal: trying another approach after a failure, adapting to new information, and asking for clarification or checking in before an important action. Meta’s newsroom post frames the target uses as expert work in law, finance and health, plus coding for games and business apps, and carries two short films in which the researchers who built the models explain them.

What Meta’s numbers show

The model card compares both models with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, every model run at maximum effort, and the results split cleanly by task. On Meta’s agent benchmarks, Wonder Full leads most rows: it scores 59.1 on Meta’s internal Personal Superintelligence Index against 55.2 for Opus 5.5, 45.4 on AutomationBench-AA against 41.6 for Astra, and 13.1 on Harvey’s legal-agent benchmark, where Astra manages 8.6 and Opus 5.5 only 4.2. On MemoryBench, the test behind Meta’s memory pitch, the smaller Muse Wonder scores 85.6 and Wonder Full scores 85.5, both ahead of Opus 5.5 at 81.5, Astra at 81.4 and Fable 5.1 at 79.0. The rivals keep some rows: Astra tops GDP.pdf at 34.2 and Opus 5.5 tops HealthBench Pro at 65.6.

Coding is more mixed. Wonder Full is ahead on the long-horizon tests, with 79.0 on DeepSWE 1.1 against 74.2 for Opus 5.5, and 57.5 on SWE-Marathon, which Meta describes as ultra-long-horizon engineering, against 46.3, although that SWE-Marathon run used the xhigh setting rather than max. On shorter, sharper tasks Opus 5.5 stays in front: 59.6 to 57.1 on Terminal-Bench 4.0-AA and a wide 70.5 to 59.0 on ProgramBench, which asks for programs written from scratch. Opus 5.5 also leads the OSWorld V2.1 computer-use test on partial credit, 81.8 to 78.4.

The independent read, and the bill

Artificial Analysis, whose Intelligence Index is the last row of Meta’s own table, puts Muse Wonder at 52 and Wonder Full at 56. That places Wonder Full above Fable 5.1 and Astra, both at 53, and behind Opus 5.5 at 58; the regular Muse Wonder, the one developers can actually call, sits just below the two 53s. Running the full index cost $4.52 per task on Muse Wonder, which used about 130 million output tokens. Opus 5.5 lists at the same $4 and $20 but spent twice the tokens and $5.98 per task; Astra, at a $10 and $50 list price, used 60 million tokens and cost $3.26 per task for one point more. Fable 5.1 cost $7.63. Wonder Full has no cost figure, since it is not on the API.

On safety, Meta calls Muse Wonder its “most aligned model to date” and says it carries its strongest safeguards so far against chemical, biological, radiological, nuclear and cyber misuse, while keeping defensive cyber abilities under its Superintelligence Scaling Framework. The models are trained to avoid sharing private information without permission, which matters more for a system whose headline feature is remembering things about its users. The full safety report is promised “soon” and was not out at launch.

Six months ago Meta’s lab was shipping its first model; Muse Spark 1.3 then made its name on price. Wonder is the first time the lab has put a model next to Anthropic’s and OpenAI’s best on its own terms, and the table it chose to publish is honest about where it falls short. The part no table can settle is the one Meta is betting on: whether people will want an assistant that remembers this much, once it lands in the Muse app.