Meta released Muse Wonder and Muse Wonder Full, calling them its most capable models for agents that handle personal and professional tasks. The lineup targets long, multistep assignments such as working through an insurance claim, with coding plus law, finance and health named as strengths. On the Harvey Legal Agent Benchmark run by Artificial Analysis, Muse Wonder Full passed 13.1% of tasks against 8.6% for GPT-6 Astra. For companies, this matters because agent quality is now measured by completed deliverables, not chat answers.
How Meta built the Wonder models
Both models were trained at the gigawatt-scale Prometheus data center, building on pretraining and reinforcement learning methods introduced with Muse Spark in April. Prompts can mix text with video, audio and images across a 1 million-token context window, while answers come back only in text. Muse Wonder takes over the top of the Muse lineup from Muse Spark 1.3, released by Meta Superintelligence Labs six weeks ago. Chief AI Officer Alexandr Wang, who runs the lab, had described that prior model to Bloomberg as competitive with Claude Fable 5.1.
Two features define the behavior. Active Memory carries context about a user from one conversation to the next and brings details up once they become relevant. Active Focus is meant to keep the model working when a task gets hard: a stuck model tries other approaches, recovers from errors, and checks in with the user before following through on anything important. Meta calls Muse Wonder its most aligned model to date. On an agentic misalignment test, Muse Wonder Full showed a 0% harmful-action rate.
Benchmark comparisons came with caveats. Meta compared Muse Wonder Full against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, with most tests run at maximum reasoning effort. Meta cautioned its runs of rival models were best-effort and prompts and tools may not have been optimally tuned for them. Coding results were split: Muse Wonder Full led DeepSWE 1.1 and SWE-Marathon, posting 57.5% on the latter against 46.3% for Claude Opus 5.5. The card notes those SWE-Marathon figures are not a controlled comparison.
What this means for companies adopting agents
Availability is tiered through Meta channels. Developers can use Muse Wonder starting today through the Meta Model API and Muse Code, the coding agent introduced in August. Muse Wonder Full is limited to Muse Code subscribers. Meta said Muse Wonder will power its Muse agent soon, with subscribers getting it first. The Muse personal agent has run on Muse Spark since its U. S. launch on Sept. 8, so the upgrade path runs through existing Meta agent products.
Safety and cyber results require close reading before deployment. Meta promised a full safety report soon, while the model card says it does not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological threats without mitigations. With safety post-training and system-level classifiers in place, the model rates moderate or lower across three risk domains. On ExploitGym, 92 flags narrowed to 21 through intended vulnerabilities, with 71 through unintended weaknesses, and only 21 counts for ranking.
The signal to watch is the promised full safety report plus wider benchmark confirmation. Opus 5.5 still led ProgramBench by 70.5% to 59%, HealthBench Pro, OSWorld V2.1 and the Artificial Analysis Intelligence Index by 58 to 56. If independent runs confirm the legal and long-horizon coding lead under controlled settings, buyers gain a stronger case for document-heavy and multistep work. If rivals close the gap, model choice will turn on access terms and safeguards.
