Meta releases Muse Wonder AI models built to run personal and professional agents

Meta releases Muse Wonder AI models built to run personal and professional agents

Meta Platforms Inc. today released Muse Wonder and Muse Wonder Full, two artificial intelligence models it says are its most capable yet at running agents that handle personal and professional tasks.

Meta aimed the release at the longer, multistep jobs people hand off to an AI agent. One example it gave was working through an insurance claim. Coding is a strength, the company says, as are specialist fields such as law, finance and health.

Active Memory, one of two new features in the models, carries context about a user from one conversation to the next, bringing details up once they become relevant. The other, Active Focus, is meant to keep the model working when a task gets hard. According to Meta’s announcement, a stuck model will try other approaches and recover from errors, and it checks in with the user before following through on anything important.

Meta trained both models at its gigawatt-scale Prometheus data center, building on pretraining and reinforcement learning methods it introduced with Muse Spark in April. Prompts can mix text with video, audio and images across a 1 million-token context window. Answers come back only in text.

In the model card, Meta compared Muse Wonder Full against OpenAI Group PBC’s GPT-6 Astra and two Anthropic PBC models, Claude Fable 5.1 and Claude Opus 5.5. Most tests ran every model at maximum reasoning effort. Meta’s runs of competing models were best-effort, the card cautions, and prompts and tools may not have been optimally tuned for them.

Muse Wonder Full delivered a passing result on 13.1% of tasks in Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, where a deliverable counts only if it meets every criterion and passes hallucination checks. GPT-6 Astra came closest among the rivals at 8.6%.

Coding results were split. Muse Wonder Full led two long-horizon tests, DeepSWE 1.1 and SWE-Marathon, and posted 57.5% on the latter to 46.3% for Claude Opus 5.5. The card notes the SWE-Marathon figures are not a controlled comparison. Codex and Claude Code produced the leaderboard scores for GPT-6 Astra and Claude Fable 5.1, and Meta’s model ran at an “xhigh” reasoning setting for that test. ProgramBench, which asks an agent to rebuild a complete program from scratch, went to Opus 5.5 by 70.5% to 59%.

Claude Opus 5.5 also finished ahead on HealthBench Pro and the OSWorld V2.1 computer-use benchmark. On the Artificial Analysis Intelligence Index, a composite of evaluations the benchmarking firm runs itself, the Anthropic model came out two points clear at 58.

Meta has promised a full safety report soon. In the meantime, the model card says Meta does not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological threats when it runs without mitigations. Mitigations aimed at that domain include safety post-training and system-level classifiers. With those safeguards in place, the model rates “moderate or lower” across all three risk domains in the Meta Superintelligence Scaling Framework, including cybersecurity and loss of control.

Of the 92 flags Muse Wonder Full captured on the ExploitGym cyber benchmark, a verification pass found that 21 came through the vulnerability each challenge was designed around. The model got the other 71 through unintended weaknesses. Because ExploitGym’s published ranking counts only the intended route, 21 is the figure to compare with it, the company said. The cyber results come two months after Meta disclosed that one of its models had breached an outside organization during a cybersecurity test in which a setup error left the sandbox connected to the internet.

Meta’s research post calls Muse Wonder “our most aligned model to date.” On an agentic misalignment test, which checks whether an agent will act harmfully to advance its goals or avoid replacement or shutdown, Muse Wonder Full’s harmful-action rate was 0%.

Muse Wonder takes over the top of the Muse lineup from Muse Spark 1.3, which Meta Superintelligence Labs released six weeks ago. Chief AI Officer Alexandr Wang, who runs the lab, described that model to Bloomberg as “competitive” with Claude Fable 5.1. Meta’s personal agent, Muse, has run on Muse Spark since its U.S. launch on Sept. 8.

Developers can use Muse Wonder starting today through the Meta Model API and Muse Code, the coding agent Meta 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.

Image: Meta