Meta releases the Muse Wonder models for personal and work agents

Meta released two models, Muse Wonder and Muse Wonder Full, which it called its most capable to date. They follow Muse Spark, the closed model from April that began the line. Muse Wonder is available to developers through the Meta Model API and the Muse Code coding tool. Muse Wonder Full, the higher-end version, is limited to Muse Code subscribers. Meta said Muse Wonder would soon power its Muse personal agent, going to paying subscribers first. The weights were not released. Both models were trained at Meta’s gigawatt-scale Prometheus data centre and accept up to a million tokens of context.

Meta built its case on agent work rather than general knowledge. Two features were named: “Active Memory”, for carrying a user’s details across conversations, and “Active Focus”, for keeping at a task through errors and checking in before consequential actions. In Meta’s comparison table, with every model at maximum reasoning effort, Muse Wonder Full led Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra on several tests. It scored 79.0% on the DeepSWE long-horizon coding benchmark, a run that Datacurve, the benchmark’s maker, carried out itself. It also led on Harvey’s legal agent benchmark and Vals’s finance agent test. It trailed on others. Opus 5.5 was ahead on OSWorld 2.1 computer use, Terminal-Bench 4.0, HealthBench Pro and the ProgramBench from-scratch coding test, where it scored 70.5 against 59.0. On the Artificial Analysis Intelligence Index, which the independent firm runs itself, Muse Wonder Full scored 56 to Opus 5.5’s 58, fourth among the models it tracks. Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens in its API preview.

The model card says where each number came from. Some are Meta’s own runs. Others are Artificial Analysis versions of third-party benchmarks, and a few are figures the rival labs reported themselves. Meta noted that its runs of competitors’ models “may not be optimally tuned”, and that the SWE-Marathon row compared models run through different agent harnesses. Several headline results, including Meta’s “Personal Superintelligence Index” and MemoryBench, come from internal benchmarks.

Under its Superintelligence Scaling Framework, Meta assessed the models’ remaining risk in cybersecurity, chemical and biological weapons, and loss of control as “moderate or lower” with safeguards in place. A fuller safety report was promised but not yet published. Coming a day ahead of Gemini 4 Argon’s general release, the launch put Meta back among the labs that publish head-to-head frontier comparisons, six months after it had stopped releasing its best models as open weights.