Meta Releases Muse Wonder AI Models for Agentic Workflows Across Law, Finance, and Health
New models introduce cross-conversation memory and error recovery for multi-step professional tasks, though early benchmarks show mixed performance against Anthropic and OpenAI competitors.
Meta Platforms Inc. has released two new AI models designed to power agents that handle personal and professional tasks across specialized domains including law, finance, and health. Released October 12, 2026, Muse Wonder and Muse Wonder Full introduce features tailored for agentic workflows—AI systems that perform multi-step tasks with minimal human intervention—though early benchmark results reveal mixed performance against competitors.
Specialized Features for Agent Workflows
Both models were trained at Meta's gigawatt-scale Prometheus data center using pretraining and reinforcement learning methods inherited from Muse Spark, released in April 2026. The architecture supports a 1 million-token context window and accepts multimodal inputs including text, video, audio, and images, though outputs remain text-only.
Two features distinguish the Muse Wonder line for professional deployment: Active Memory, which retains context across separate conversations to improve continuity in client encounters or administrative workflows, and Active Focus, an error recovery system that tries alternative approaches when tasks become difficult and prompts users to verify critical actions before execution. According to Alexandr Wang, Chief AI Officer at Meta Superintelligence Labs, the models target longer, multistep jobs such as working through insurance claims.
Mixed Performance Across Professional Benchmarks
Meta's models demonstrated strong results in legal and coding tasks—achieving a 13.1% pass rate on Harvey's Legal Agent Benchmark compared to GPT-6 Astra's 8.6%, and scoring 57.5% on the SWE-Marathon coding test versus Claude Opus 5.5's 46.3%. However, performance varied across other domains. On HealthBench Pro, Anthropic's Claude Opus 5.5 outperformed Muse Wonder Full, though Meta did not disclose specific scores. Claude Opus 5.5 also led on the OSWorld V2.1 computer-use benchmark and the Artificial Analysis Intelligence Index composite, scoring 58 compared to Muse Wonder Full's 56.
In coding benchmarks, results were split. While Muse Wonder Full led on DeepSWE 1.1 and SWE-Marathon, Claude Opus 5.5 won on ProgramBench with 70.5% versus 59%. Meta's model card cautions that competing models were run on a best-effort basis and prompts may not have been optimally tuned for them.
Safety Profile and Deployment Considerations
Meta's safety assessments revealed a nuanced risk landscape for professional applications. The company does not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological threats when running without mitigations. With safeguards in place—including safety post-training and system-level classifiers—Meta rates the models at "moderate or lower" risk across all three domains in its Superintelligence Scaling Framework, including cybersecurity and loss of control.
The models achieved a 0% harmful-action rate on agentic misalignment testing, addressing concerns about AI systems taking unauthorized actions to advance their goals or avoid replacement. Meta's research post calls Muse Wonder "our most aligned model to date."
On the ExploitGym cybersecurity benchmark, verification found that 21 of 92 flags came through the vulnerability each challenge was designed around, while the model captured the other 71 through unintended weaknesses. Because ExploitGym's published ranking counts only the intended route, Meta noted that 21 is the figure for comparison. These 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.
Availability and Market Positioning
Muse Wonder is available through Meta's Model API and will power Muse Code, the company's coding agent introduced in August. Muse Wonder Full remains limited to Muse Code subscribers. The models will eventually power Meta's Muse personal agent, with subscriber access prioritized. The release replaces Muse Spark 1.3, released just six weeks earlier in September 2026, signaling Meta's accelerated development timeline in the competitive AI agent market.
For developers working in specialist fields, the models offer a new enterprise option with demonstrated cross-domain capabilities, though the benchmark results indicate careful evaluation against existing solutions remains essential before production deployment.
Sources
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.