Meta introduced the Muse Wonder series on October 12, presenting Muse Wonder and Muse Wonder Full as “frontier models” for personal and professional agentic tasks. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Meta also plans to bring Muse Wonder to its Muse personal agent, initially for subscribers.
Memory and long-running tasks
Meta claims the models deliver “state-of-the-art” results across benchmarks covering agents, memory, and coding. The models are intended to handle work involving leases, financial information, health information, law, finance, and health.
The company highlights two features for multi-step tasks:
- Active Memory carries context across conversations and surfaces it when relevant.
- Active Focus keeps working toward a goal when a task becomes difficult, including trying alternative approaches and working through blockers.
On MemoryBench, which measures how models retain and use context across a long interaction history, Muse Wonder scored 85.6% and Muse Wonder Full scored 85.5%. Those results were above the listed scores for Claude Opus 5.5 at 81.5%, GPT-6 Astra at 81.4%, and Claude Fable 5.1 at 79%.
Meta’s internal FocusBench evaluation, which involves realistic multi-step tasks with tool use and file processing, produced a more mixed comparison between the two new models. Muse Wonder Full scored 54.2%, ahead of GPT-6 Astra at 47.7% and Claude Opus 5.5 at 45.8%. Muse Wonder scored 44.9%, while Claude Fable 5.1 scored 36.6%.
Results vary by benchmark
The reported results do not establish a clean lead across every test. Muse Wonder Full scored 59.1 on the Personal Superintelligence Index, compared with 55.2 for Claude Opus 5.5, 54.5 for GPT-6 Astra, and 49.9 for Claude Fable 5.1.
On HealthBench Pro, however, Muse Wonder Full scored 62.3 and Muse Wonder scored 60.5, below GPT-6 Astra at 65.3 and Claude Fable 5.1 at 65.6. On Terminal-Bench 4.0-AA, Muse Wonder Full scored 57.1, behind Claude Opus 5.5 and Claude Fable 5.1 at 59.6 and GPT-6 Astra at 59.1.
The same pattern appears in coding evaluations. Muse Wonder Full scored 79.0 on DeepSWE 1.1 and 57.5 on SWE-Marathon, while Muse Wonder scored 78.8 and 53.1, respectively. On ProgramBench, Muse Wonder Full scored 59.0 and Muse Wonder scored 50.0, behind Claude Fable 5.1 at 70.5.
Meta also reports scores of 78.4 on the partial metric and 39.4 on the binary metric for Muse Wonder Full on OSWorld V2.1. Muse Wonder scored 72.6 and 34.3 on those metrics.
Safety and availability
Meta claims the Muse Wonder family was trained to resist adversarial instructions and handle private information carefully. On the Cyber Misuse Refusal Benchmark, Muse Wonder Full recorded a 96.9% refusal rate and Muse Wonder recorded 93.7%. Claude Fable 5.1 scored 97.5%, while GPT-6 Astra scored 86.3%.
Muse Wonder is available now through the Meta Model API and Muse Code. Muse Wonder Full is limited to Muse Code subscribers, and Meta has not provided further availability details in the announcement. The company also mentioned that open weights for Muse Spark are “coming soon.”
More details are available in Meta’s research blog.
Source: AI at Meta on X
